Quality detection system and method for glue injection laser profile

Abstract

The application discloses a quality detection system and method for glue injection laser profile, and belongs to the technical field of quality detection. The system comprises the following steps: obtaining a reflected point cloud signal of a glue injection finished product, constructing a point cloud data stream, setting an image reconstruction method, constructing a glue layer three-dimensional model, synchronously generating a pseudo-color gray scale image, and realizing visual representation of the surface texture of the glue layer; constructing a three-layer feature extractor, setting a quality preliminary judgment method, cross-layer fusing features of different scales, solving the missed detection problem of traditional single feature detection on complex defects, automatically identifying basic defects, and outputting a quality preliminary judgment result; judging defects of the glue injection finished product that does not meet the preliminary judgment, setting a defect detailed judgment method, voxel-level segmenting a three-dimensional image, calculating features of the defects, and judging whether the defects affect the use performance of the product, so as to generate a detailed judgment result, realize accurate positioning of complex defects, and improve the defect type recognition accuracy.

Description

Technical Field

[0001] This invention relates to a quality inspection system and method for laser contouring of adhesive injection, belonging to the field of quality inspection technology. Background Technology

[0002] In industrial glue injection production, the quality of glue injection directly affects product performance. Existing detection methods suffer from insufficient accuracy and low automation. Laser contour detection technology, with its non-contact measurement advantage, can actively emit lasers and receive reflected signals through laser sensors to acquire three-dimensional data of the glue layer in real time. Combined with image processing algorithms, it can achieve glue layer contour fitting and quantitative detection of regular defects, improving detection accuracy and efficiency, and meeting the high-precision and automated requirements of smart manufacturing for glue injection quality inspection.

[0003] However, due to the interaction between laser signal transmission, reflection characteristics, and complex industrial environments, existing technologies still face multiple challenges. For example, highly reflective substrates cause specular or diffuse reflection of the laser, resulting in energy attenuation and noise superposition at the receiving end, which damages the integrity of point cloud data. Furthermore, the transparent adhesive layer generates dual reflection signals from the substrate and adhesive layer surfaces due to the laser penetration effect, leading to confusion in contour data. Environmental pollutants such as dust and oil further scatter the laser beam, causing a decrease in the signal-to-noise ratio of the received signal, which significantly reduces the boundary segmentation accuracy between the adhesive layer and the substrate. At the same time, in defect identification, the judgment of regular defects by manually setting fixed thresholds can only cover standard defects such as glue width exceeding tolerance and glue height being insufficient. It is not capable of extracting features of non-standard defects such as glue climbing, glue collapse, and bubbles, as well as various forms of adhesive materials such as high-elasticity silicone and fluorescent glue. Moreover, it requires manual preset detection parameters, making it difficult to adapt to the complex and ever-changing quality inspection needs in new processes. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a quality inspection system and method for glue injection laser contouring. It constructs a three-dimensional contour of the glue injection area using line laser scanning technology, achieves high-precision quality judgment by combining image processing algorithms, and overcomes the signal interference bottleneck in complex environments by combining intelligent defect identification methods, thereby improving the system's ability to detect non-standard defects and multi-shaped glue materials.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] The quality inspection system for laser contouring of glue injection includes:

[0007] Acquire the reflected point cloud signal of the glued product, construct the point cloud data stream, set the image reconstruction method, construct the three-dimensional model of the glue layer, and simultaneously generate a pseudo-color grayscale image;

[0008] A three-layer feature extractor is constructed, a quality assessment method is set, features of different scales are fused across layers, basic defects are automatically identified, and the quality assessment results are output.

[0009] For finished glued products that initially fail the quality assessment, a defect judgment method is set up to further judge the defects. The three-dimensional image is segmented at the voxel level to calculate the characteristics of the defects and determine whether the defects affect the performance of the product, so as to generate the detailed judgment results.

[0010] Specifically, the image reconstruction method includes:

[0011] The initial point cloud data stream is converted into a unified format, invalid points are removed, and a point cloud dataset is formed. In the formula, a represents the number of point clouds when a complete scan is completed, and p i Let i be the i-th point in the point cloud dataset;

[0012] For point p i Select the point p i Find the points with the smallest Euclidean distance between them to define a local neighborhood;

[0013] Obtain the maximum and minimum values ​​of the reflection intensity within the local neighborhood, and calculate the value of point p through linear transformation. i The grayscale value g(i) is obtained and saved to the cloud dataset mentioned above;

[0014] Sort the gray values ​​within the local neighborhood in ascending order and calculate the median gray value g within the local neighborhood. med and standard deviation σ g ;

[0015] Filter out |g(i)-g from the point cloud dataset med |>kσ g The points are then removed to form an effective contour point cloud.

[0016] Specifically, the image reconstruction method further includes:

[0017] Set the interior point threshold and the maximum number of iterations, and randomly select three points from the effective contour point cloud to calculate the plane equation and obtain the initial fitting plane;

[0018] Calculate the distance from points in the effective contour point cloud to the fitting plane, retain points whose distance does not exceed the inlier threshold, and define them as inliers;

[0019] The calculation is iterated continuously until the maximum number of iterations is reached, thus obtaining the optimal plane and its interior point set.

[0020] When the transparent adhesive layer is detected, the thickness of the adhesive layer is calculated according to the Beer-Lambert law;

[0021] Thickness compensation is used to compensate for the intensity of the reflected signal penetrating to the substrate, and the original intensity in the inner point set is replaced with the compensated reflection intensity to generate a pure point cloud.

[0022] Set the sphere radius, generate a triangular mesh, define the initial control matrix and node vectors, optimize the surface parameters using the least squares method, and generate a parameterized surface model S(θ,ξ); where θ and ξ are surface parameters.

[0023] The grayscale values ​​in the prime pure point cloud are mapped to the surface of the triangular mesh through vertex attributes to generate a grayscale image containing surface texture information.

[0024] Specifically, the initial quality assessment method includes:

[0025] The vertex coordinates of the parameterized surface model are normalized to generate a geometric tensor, and the grayscale image is bilinearly interpolated to generate a grayscale tensor.

[0026] Calculate the gradient and edge intensity of the vertices in the x-axis, y-axis, and z-axis directions in the geometric tensor to obtain the edge tensor;

[0027] Each point in the geometric tensor is mapped to the corresponding voxel, and the point cloud information in each voxel is statistically analyzed to form voxelized 3D data.

[0028] Local features and spatial structure information are extracted from the voxelized 3D data to obtain a feature map. Specific geometric features are then extracted from the feature map, normalized, and a geometric feature vector is formed.

[0029] The grayscale tensor is processed to generate a multi-scale feature map, and a global average pooling operation is performed to generate a texture feature vector.

[0030] Specifically, the initial quality assessment method further includes:

[0031] For each point in the geometric tensor Calculation points The Euclidean distance between the point and other points in the geometric tensor is selected. The nearest l points form a neighborhood set;

[0032] Substitute the coordinates of the points in the neighborhood set into the constructed plane equations, and use least squares plane fitting to obtain an overdetermined linear system of equations.

[0033] Repeat the neighborhood selection and least squares plane fitting steps described above until all points in the geometric tensor are traversed, resulting in the set of normal vectors of the point cloud, which together form the normal vector tensor.

[0034] Based on the spatial distribution of point clouds in the parametric surface model, the relationship between adjacent points is defined, the angle between adjacent normal vectors is calculated, and points with angles greater than a preset threshold are marked as significant contour points; otherwise, they are marked as non-contour points, and a mask vector is generated. For the contour points, let M χ =1, for non-contour points, let M = 1. χ =0;

[0035] The weight coefficient of the contour points is configured as w M.1 The weighting coefficient for non-contour points is w. M.2 Canny detection is used to identify edge regions of the grayscale image, and the weight coefficient of the edge regions is configured as w. g The geometric feature vector and the texture feature vector are fused using weighting coefficients to generate a fused feature vector.

[0036] Specifically, the initial quality assessment method further includes:

[0037] Set initial parameters θ0, train a meta-learning model using the MAML algorithm, collect samples of the new adhesive material, extract fusion feature vectors, update model parameters according to the meta-learning algorithm, and obtain model parameters θ' suitable for the new adhesive material;

[0038] The process parameters of glue injection are collected in real time, and the process parameters and historical glue layer height data are input into the constructed correlation model to output the predicted glue layer height under the current process parameters.

[0039] Calculate the deviation pressure between the current injection pressure and the standard pressure. If the absolute value of the deviation pressure does not exceed the preset deviation threshold, the process parameters have not fluctuated; otherwise, the process parameters have fluctuated, and the qualified glue height threshold is dynamically adjusted to output the allowable range of glue layer height.

[0040] The student model is used to load the adapted model parameters θ' to obtain the fusion feature vector of the glue-injected finished product, and then the multi-class probability distribution is output through a fully connected layer.

[0041] If the predicted probability of the defect category exceeds the preset defect threshold and the adhesive layer height exceeds the allowable range, the product is judged as unqualified; otherwise, it is judged as qualified.

[0042] The output quality assessment results include the quality grade of the finished glue-filled product and the corresponding preliminary defect labels.

[0043] Specifically, the defect analysis method includes:

[0044] Obtain data on defective products and extract geometric feature thresholds based on the initial defect labels;

[0045] Based on the geometric feature threshold, suspicious regions corresponding to different defect types are divided, and sub-images corresponding to the suspicious regions are cropped from the grayscale image. The point cloud data of the suspected defect regions and the corresponding grayscale image regions are output.

[0046] Voxelization and position encoding are performed on the point cloud data of suspected defect areas to generate voxel-level 3D defect segmentation masks.

[0047] Features are captured using a multi-head self-attention mechanism, and after fusing Transformer and U-Net++ features, defect boundaries are generated using CRF.

[0048] Using the defect boundary points as nodes, nodes with an Euclidean distance less than a preset threshold are connected, and the cosine of the angle between the node normal vectors is used as the edge weight to construct the defect graph structure.

[0049] The feature information of nodes and their neighbors is aggregated and calculated to generate graph-level feature vectors, which are then concatenated with the geometric features to generate multimodal fusion feature vectors.

[0050] Based on the multimodal fusion feature vector and real-time glue injection process parameters, and combined with the rules in the preset knowledge graph, the unqualified products are classified into first-level defects and second-level defects. First-level defects are eliminated, and second-level defects are re-injected. The decision rules of the knowledge graph are periodically optimized to output the correction results.

[0051] Specifically, the system also includes:

[0052] Read the point cloud data stream, set the adaptive configuration method, obtain environmental parameters, monitor environmental parameters in real time, and dynamically adjust the laser emission power.

[0053] Specifically, the adaptive configuration method includes:

[0054] Environmental parameters are collected by sensors placed at the outlet of the dispensing area, and an environmental parameter vector is generated by median filtering.

[0055] Based on the substrate type, the base power is set, the environmental parameters are normalized, and the impact factor is calculated.

[0056] The power regulation rate is determined by combining the fuzzy rule base, and the actual transmission power is calculated.

[0057] Hardware triggering and timestamp matching techniques are used to synchronize point cloud and environmental data, interpolate mismatched data, and generate structured point cloud data.

[0058] In the detection of substrates with reflectivity exceeding a preset value, a polarizer is installed, and the transparent adhesive layer uses pulse code modulation and multispectral technology. A dust compensation model is established based on Mie scattering theory to correct the signal.

[0059] Quality inspection methods for laser-guided adhesive injection contours include:

[0060] A laser sensor is used to emit a laser to the finished glue-filled product and receive the laser signal. Environmental parameters are monitored in real time, the laser emission power is dynamically adjusted, and structured point cloud data is generated.

[0061] The point cloud data is converted into a new format and invalid points are removed. The point cloud data is then fitted to generate a parametric surface model and a grayscale image.

[0062] The extracted geometric and texture features are fused together, and a meta-learning model is used to quickly adapt to new adhesive materials. The qualified threshold is dynamically adjusted by combining the correlation model, and the initial quality judgment result is output.

[0063] For finished glued products that initially fail the inspection, a defect assessment is conducted to generate detailed assessment results.

[0064] The beneficial effects of this invention are:

[0065] By integrating multiple technologies, the challenges of adhesive injection testing for complex materials are effectively addressed. For highly reflective substrates and transparent adhesive layers, environmental parameters are collected in real time by sensors. Combined with fuzzy rules, laser power is dynamically adjusted. Polarizers are added, and pulse coding and multispectral technologies are employed to resolve the issues of specular reflection blind spots and signal penetration confusion, thereby improving signal quality. Image reconstruction technology removes noise and compensates for signal attenuation, generating a high-precision 3D model that provides a visual basis for quality judgment. Initial quality judgment and detailed defect judgment are combined with meta-learning, LSTM, and knowledge graphs to achieve multi-scale feature fusion and precise defect localization, avoiding missed detections and subjective errors. Based on the results, strategies are generated to link equipment, forming a closed loop between testing and production. This adapts to process fluctuations, significantly improving the surface inspection accuracy of complex materials and the automation level of the production line, while reducing the defect rate. Attached Figure Description

[0066] Figure 1 A structural diagram of a quality inspection system for laser contouring of adhesive injection.

[0067] Figure 2 This is a flowchart of the image reconstruction method of the present invention;

[0068] Figure 3 This is a flowchart of the initial quality assessment method of the present invention;

[0069] Figure 4 This is a flowchart of the defect analysis method of the present invention;

[0070] Figure 5 Flowchart of the quality inspection method for laser contouring in glue injection. Detailed Implementation

[0071] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0072] Example 1

[0073] refer to Figures 1 to 4 As shown in the figure, this embodiment introduces a quality inspection system for laser contouring of glue injection, including: a sensing module, a processing module, a quality judgment module, an identification module, and an execution module;

[0074] The sensing module is used to emit structured light to the glued finished product using a line laser sensor, receive the reflected laser point cloud signal in real time, and form an initial point cloud data stream containing distance and intensity information. It sets an adaptive configuration method, acquires environmental parameters, monitors environmental interference parameters in real time, and dynamically adjusts the laser emission power to avoid confusion between the specular reflection blind zone of the highly reflective substrate and the signal penetration of the transparent adhesive layer.

[0075] The processing module is used to preprocess cloud data streams, set image reconstruction methods, fit point cloud data, construct a three-dimensional model of the adhesive layer, fully restore the outline of the adhesive layer, including height, width, and surface curvature, and simultaneously generate pseudo-color grayscale images, which are superimposed on the three-dimensional model of the adhesive layer to achieve a visual representation of the surface texture of the adhesive layer, thereby generating a high-precision three-dimensional image of the adhesive layer outline, providing a visual and quantitative basis for subsequent quality judgment.

[0076] The quality assessment module is used to construct a three-layer feature extractor, set the initial quality assessment method, and achieve cross-layer fusion of features at different scales through the feature pyramid network to solve the problem of missed detection of complex defects by traditional single feature detection. The module pre-trains a lightweight ResNet-18 model and achieves rapid adaptation for different adhesives and substrates by fine-tuning the last three fully connected layers. It automatically identifies basic defects and outputs the initial quality assessment results.

[0077] The identification module is used to judge defects in the glued finished products that are initially deemed unqualified. It sets up a detailed defect judgment method, performs voxel-level segmentation on the 3D image, distinguishes the glue layer, substrate and defect area, outputs the spatial coordinates and volume parameters of the defect, calculates the characteristics of the defect by combining the gray-level co-occurrence matrix, analyzes the texture consistency of the defect area, and judges whether the defect affects the performance of the product. It outputs the detailed judgment results to achieve accurate positioning of complex defects, improve the accuracy of defect type recognition, and avoid the subjective error of human experience judgment.

[0078] The execution module is used to generate differentiated processing strategies based on the initial quality assessment results and the detailed defect assessment results, and to link with external equipment to achieve closed-loop control of inspection and production, thereby improving the automation level and yield of the production line.

[0079] In this embodiment, the glue injection quality inspection takes a toothbrush as an example. The glue injection location is the brush head base. By injecting glue, the roots of the bristles are fixed in the grooves of the brush head base to ensure the connection strength between the bristles and the brush handle, so as to prevent the bristles from falling off. At the same time, glue is injected in the transition connection area between the brush head and the brush handle to enhance the bonding strength between the brush head and the brush handle, so as to avoid the brush head breaking due to force (such as bending when brushing teeth) during use. Some toothbrushes also have an anti-slip or decorative layer. The anti-slip rubber layer is bonded to the brush handle body by glue injection, or an adhesive layer with concave and convex texture is formed on the surface of the brush handle.

[0080] Specifically, image reconstruction methods include:

[0081] The initial point cloud data stream is converted into a unified format, such as the PCD point cloud data format, and invalid points are removed to form a point cloud dataset P = {p1, ..., p...}. a}, where a is the number of point clouds when a complete scan is completed, p i Let p be the i-th point in the point cloud dataset, i = 1, ..., a, p i =(x i ,y i ,z i ,I i ,t i ), (x i ,y i ,z i Let p be a point. i The coordinates of I i For point p i The reflection intensity, t i For point p i timestamp;

[0082] To enhance data visualization and adapt to subsequent grayscale-based image processing algorithms, the reflection intensity in the point cloud dataset is mapped to 8-bit grayscale values. A local normalization method is used for each point p. i The local neighborhood is defined, containing several points, such as the 50 points with the closest Euclidean distance. The maximum and minimum values ​​of the reflection intensity are obtained, normalized by linear transformation, and the calculated gray values ​​are saved to the point cloud dataset P. i =(x i ,y i ,z i ,I i ,g(i),t i ), g(i) is a point p iThe grayscale values ​​not only preserve the relative differences in reflection intensity, but also convert continuous signals into discrete grayscale values, providing a unified measurement space for subsequent noise detection;

[0083]

[0084] In the formula, I max I min ∈ represents the maximum and minimum values ​​of the reflection intensity within the local neighborhood, and ∈ represents a minimum value to avoid a denominator of zero;

[0085] Noise points are identified using a combination of median filtering and standard deviation. For each point p... i Sort the gray values ​​in the local neighborhood in ascending order, g(1)≤g(2)≤…≤g(b), and calculate the median gray value g in the local neighborhood. med and standard deviation σ g Where b is the number of gray values ​​in the local neighborhood, and when b is odd, the median gray value is g. med.1 When b is even, the median gray value is g. med.2 The expression is as follows:

[0086]

[0087]

[0088] Filter out |g(i)-g from the point cloud dataset med |>kσ g Points are defined as noise points and removed, while edge contour points are effectively retained and random noise points are filtered out to form an effective contour point cloud P1; where k is a screening coefficient, and in this embodiment, k = 2.5;

[0089] For adhesive layer detection scenarios, a random sampling consensus algorithm is used to separate planar substrates from non-planar adhesive layer edges. An in-point threshold and a maximum number of iterations are set. Three points are randomly selected from the effective contour point cloud, and these three points are not collinear. The plane equation is calculated using the selected three points to obtain the initial fitting plane. The distance from all points in the effective contour point cloud to the plane is calculated, and points whose distance does not exceed the in-point threshold are retained and defined as in-points.

[0090] The process involves iterative calculations, including random point selection, plane fitting, and internal point filtering, until the maximum number of iterations is reached. This yields the optimal plane and internal point set P2, effectively eliminating abnormal points caused by specular reflection, such as strong light reflection points on smooth metal surfaces, thus providing a clean substrate point cloud for subsequent adhesive layer thickness compensation.

[0091] To correct for the attenuation of the reflected signal as the laser penetrates the adhesive layer to the substrate and to restore the true adhesive layer thickness, when the transparent adhesive layer is detected, the adhesive layer thickness is calculated according to the Beer-Lambert law, while the substrate reflection intensity I is collected. base Because the adhesive layer attenuation is underestimated, thickness compensation is used to restore the true strength. This compensates for the intensity of the reflected signal penetrating to the substrate, correcting the signal attenuation caused by changes in adhesive layer thickness. This ensures that the reflection intensity values ​​in the dataset reflect the true reflection characteristics of the substrate. The compensated reflection intensity replaces the original intensity in the dataset, generating a clean point cloud P. clean The detection of a transparent adhesive layer is triggered by material marking. This is achieved through an integrated material identification sensor (such as a spectrometer or capacitive sensor) or a material label in a pre-defined workpiece CAD model. When a pre-defined adhesive layer material ID is detected, such as the material code corresponding to epoxy resin, the presence of a transparent adhesive layer in the current area is determined. The expression is as follows:

[0092]

[0093]

[0094] In the formula, d is the thickness of the adhesive layer, and I tran I represents the transmitted light intensity of the transparent adhesive layer, indicating the light intensity signal that passes through the finished adhesive layer and is collected in real time by a detector. in Let I be the initial laser intensity, μ be the material absorption coefficient, and I be the initial laser intensity. comp The compensated reflection intensity value, d std For the preset standard thickness, ln(·) represents the logarithmic function;

[0095] Surface reconstruction is performed using the sphere-pivot algorithm. A sphere radius is set to ensure detail preservation at the adhesive layer edges. A triangular mesh is generated, and an initial control matrix V and node vector Ξ are defined. Surface parameters are optimized using the least squares method to generate a parameterized surface model S(θ,ξ). Points are selected at fixed intervals within the triangular mesh to ensure no loss of surface details. Sampling is more frequent in areas with drastic curvature changes (such as adhesive layer edges and sharp corners) and less frequent in flat areas, balancing fitting accuracy and computational efficiency. The 3D coordinates of the sampled points are arranged sequentially to form a control vertex matrix Θ = [θ1,…,θ]. c ] T θ c =(x c ,y c ,z c Let be the coordinates of the c-th control vertex. In flat regions, nodes are evenly spaced, while in regions with drastic curvature changes, the node spacing is proportional to the geometric distance between sampling points, thus generating the node vector Ξ = [ξ1,…,ξ]. d1], where θ and ξ are surface parameters, and d1 is the number of types of parameter ξ. By continuously adjusting the surface parameters, the sum of squared errors between the fitted NURBS surface equation value and the original data points is minimized. The optimal combination of surface parameters is found through iterative calculation. The expression for the sum of squared errors B is as follows:

[0096]

[0097] In the formula, G is the number of points in the pure point cloud;

[0098] The grayscale values ​​in the pure point cloud are mapped to the surface of the triangular mesh through vertex attributes, generating a grayscale image I containing surface texture information. gray For example, high-reflectivity areas appear white, while transparent adhesive layers appear semi-transparent gray, which is used for texture feature extraction in subsequent defect identification.

[0099] Specifically, the methods for initial quality assessment include:

[0100] The vertex coordinates in the parametric surface model are normalized to obtain the extreme values ​​on each coordinate axis. The original coordinate values ​​are scaled to [-1, 1] to eliminate dimensional differences and generate a geometric tensor, which helps in learning data features. At the same time, bilinear interpolation is used to adjust the grayscale image of arbitrary size from its original size to 224×224 pixels to adapt to the input requirements of subsequent CNNs, maintain edge sharpness, and generate a grayscale tensor. If the original image is RGB three-channel, the luminance channel is taken to convert it into a single-channel grayscale tensor to meet the input requirements of subsequent two-dimensional texture feature extraction. Bilinear interpolation calculates the grayscale value of the target pixel by weighted averaging of the grayscale values ​​of the four neighboring pixels around the target pixel, thus ensuring the quality of the image after scaling.

[0101] Based on the geometric tensor, the gradient of each vertex in three directions is calculated, and the edge intensity of each vertex is calculated to obtain the edge tensor, which reflects the sharpness of the adhesive layer profile and edge information.

[0102] Traversing each point in the normalized geometric tensor Each point is mapped to a corresponding voxel, and the point cloud information within each voxel is statistically analyzed, such as the number of points and the centroid, to form voxelized 3D data. This converts continuous point cloud data into a discrete voxel mesh. CNN feature extraction is then performed on the voxelized 3D data, using three layers of 3D convolution to extract local features and spatial structure information from the voxelized point cloud data. After three layers of convolution, the final feature map is obtained, and specific geometric features, such as volume, height variance, and curvature distribution, are extracted from the feature map. These geometric features are then normalized to ensure their values ​​are within a suitable range. Finally, the normalized features are concatenated into a vector in a specific order to form the final geometric feature vector F.3D Where Q is the number of midpoints in the geometric tensor. Let x be the coordinates of the x-th point in the geometric tensor, where x = 1, ..., Q;

[0103] The grayscale tensor is processed using a ResNet-18 network to generate multi-scale feature maps. These feature maps at different scales capture texture information at different levels of the image. Global average pooling is then performed on the feature maps output from the last layer of the ResNet-18 network to convert the multi-scale feature maps into texture feature vectors F. 2D It contains the texture information of the image;

[0104] The normal vector of each point in the geometric tensor is calculated using the local curve fitting method. Calculation points The Euclidean distance between the point and other points in the geometric tensor is chosen to be the distance from the point. The nearest l points form a neighborhood set Substituting the coordinates of points in the neighborhood set into the constructed plane equations, and using least-squares plane fitting, an overdetermined system of linear equations is obtained. For all points in the geometric tensor, the neighborhood selection and least-squares plane fitting steps are repeated to obtain the set of normal vectors of the point cloud. The normal vector tensor is composed of the spatial distribution of the point cloud in the parametric surface model, defining the relationship between adjacent points. For each point... Consider the l1 points that are closest to it in Euclidean distance in space as adjacent points. Calculate the angle between adjacent normal vectors. Mark points with an angle greater than a preset angle threshold as significant contour points, otherwise mark them as non-contour points. Create a binary mask vector of length Q. M χ For contour points ∈{0,1}, let M χ =1, for non-contour points, let M = 1. χ =0, to clearly identify the contour regions in the point cloud, providing a key basis for subsequent feature weighting and fusion;

[0105] Weights are assigned to the geometric feature vector and the texture feature vector respectively. For the geometric features, the weight coefficient of the contour points is set to w. M.1 The weighting coefficient for non-contour points is w. M.2 For texture features, Canny detection is used to determine the edge regions of the grayscale image, and the weight coefficient of the edge regions is configured as w. g By using weighting coefficients to fuse geometric feature vectors and texture feature vectors to generate a fused feature vector, the feature vector focuses on the features of the contour region. When detecting the injection of highly reflective or transparent materials, it effectively avoids the contour features being submerged by background information and improves the sensitivity to edge defect detection.

[0106] New adhesive materials or defect types are constantly emerging. Traditional model training methods often require a large number of samples and a long time. This paper uses the MAML algorithm to construct a meta-learning model, setting the initial parameter to θ0, to learn the general feature transformation pattern across materials and defects, such as the mapping relationship between the surface reflectivity and thickness of the adhesive layer. Historical adhesive injection datasets are collected to train the cloud learning model to quickly adapt to new tasks. During the training process, the initial parameters of the model are continuously optimized, so that when facing new tasks, the model can achieve good performance with only a small number of samples and a few gradient updates. Among them, the historical adhesive injection data is the historical fusion feature vector.

[0107] For several samples of new adhesive materials (such as fluorescent adhesive), a fusion feature vector is extracted, a task-specific loss function is calculated using these samples, the gradient of the loss function with respect to the model parameters is calculated through backpropagation, and the model parameters are updated according to the meta-learning algorithm to obtain the model parameters θ' suitable for the new adhesive material.

[0108] The process parameters of adhesive dispensing are collected in real time by the equipment's sensors. The process parameters and historical adhesive layer height data are input into the constructed association model, which outputs the predicted adhesive layer height under the current process parameters. The association model is constructed by an LSTM network to process long-sequence data and remember long-term dependency information. The association model is trained with a large amount of historical data to learn the relationship between process parameters and adhesive layer height. By continuously adjusting the model parameters, the model can accurately predict the adhesive layer height under different process parameters.

[0109] Calculate the deviation pressure between the current injection pressure and the standard pressure. If the absolute value of the deviation pressure does not exceed the preset deviation threshold, it indicates that the process parameters have not fluctuated; otherwise, it indicates that the process parameters have fluctuated significantly. Dynamically adjust the qualified adhesive height threshold to output the allowable range of adhesive layer height, so that the threshold can adapt to process fluctuations and avoid misjudgments caused by changes in process parameters; the expression is as follows:

[0110]

[0111]

[0112] In the formula, H low H high These represent the lower and upper limits of the adhesive layer height. δ is an adjustment coefficient used to predict the adhesive layer height;

[0113] The student model is used to load the adapted model parameters θ' to obtain the fusion feature vector of the current glue injection product. Several probability distributions are output through the fully connected layer, such as qualified, glue width exceeding the standard, glue height insufficient, glue breakage, and other defects. The quality level is determined based on the probability distribution and dynamic threshold.

[0114] If the predicted probability of the defect category exceeds the preset defect threshold and the adhesive layer height exceeds the allowable range, the product is judged as unqualified; otherwise, it is judged as qualified.

[0115] Output the quality grade of the finished glued product, such as qualified or unqualified, and the corresponding preliminary defect label, such as insufficient glue level or suspected glue crawling.

[0116] Specifically, the methods for detailed defect assessment include:

[0117] Data for products deemed unqualified includes parametric surface models, grayscale images, initial defect labels, and theoretical contour data of the glue-filled finished product model. The theoretical contour data of the glue-filled finished product model includes the substrate boundary coordinates, which are obtained from the glue-filled finished product design database. Precise location processing is performed on the suspected defect areas in the initial judgment. Based on the initial defect labels, the geometric feature thresholds of the corresponding defects are extracted from the process standard library and design parameters. For example, the standard height of the Z-axis is extracted for labels with insufficient glue height, and a deviation threshold exceeding the boundary of the glue-filled finished product model is set for suspected glue crawling.

[0118] Dynamic ROI delineation is performed based on geometric feature thresholds to divide suspicious areas corresponding to different defect types. In the case of insufficient glue height, the Z-axis coordinate of each vertex is calculated, and the vertex coordinates of the parametric surface model are filtered. Points with Z-axis coordinates lower than 80% of the standard height are selected and designated as suspicious areas of insufficient glue height. In the case of suspected glue climbing, the shortest distance from each vertex to the theoretical contour of the glue-filled finished model is calculated, and vertices that exceed the deviation threshold are marked and designated as suspicious areas of glue climbing.

[0119] Crops out the sub-image corresponding to the suspicious area from the grayscale image, and outputs the point cloud data of the suspected defect area and the corresponding grayscale image area;

[0120] With 0.2mm 3 The resolution is used to voxelize the point cloud data of the suspected defect area to generate a three-dimensional voxel mesh, which discretizes the continuous point cloud data to facilitate subsequent processing. The voxel mesh is then unfolded into a one-dimensional sequence, and the position code is calculated by sine and cosine functions to assign relative position information in three-dimensional space to each voxel, so as to perceive the positional relationship of voxels in space, thereby better understanding the global spatial structure of the defect and generating a voxel-level three-dimensional defect segmentation mask to mark whether each voxel is a defect.

[0121] By utilizing a multi-head self-attention mechanism, the Query, Key, and Value matrices are calculated, and long-distance dependencies across voxels are captured through a scaling dot product attention formula, such as identifying density differences between bubble defects and surrounding adhesive layers.

[0122] The Transformer output features are fused with the local features extracted by U-Net++, and then input into a Conditional Random Field (CRF). The CRF optimizes the segmentation results through an energy function, penalizes the discontinuity of adjacent voxel labels, and generates a set of defect boundary coordinates, providing accurate defect location information for subsequent feature modeling.

[0123] Using the defect boundary points as nodes, the nearest neighbor interpolation method is used to assign attribute information including three-dimensional coordinates and corresponding edge strength to each node. The Euclidean distance between nodes is calculated, and nodes with a distance less than a preset threshold are connected. The cosine value of the angle between the node normal vectors is used as the edge weight to reflect the continuity of the surface and complete the construction of the defect graph structure.

[0124] By using a multi-layer graph convolutional network, the feature information of nodes and their neighbors is aggregated and calculated to generate graph-level feature vectors to capture the overall geometric features of defects. The graph-level feature vectors are then concatenated with traditional geometric features (such as defect volume, average height, and standard deviation of curvature) to generate multimodal fusion feature vectors, providing rich feature information for the final determination of defect types.

[0125] Define entities in the knowledge graph, including defect types (such as glue creep and glue collapse), process parameters (such as injection pressure and temperature), and performance indicators (such as sealing performance and strength). At the same time, determine the relationships between these entities, such as excessive pressure causing glue creep and glue collapse, which leads to a decrease in strength. Store these relationships in a knowledge network of triples.

[0126] Based on the multimodal fusion feature vector and real-time dispensing process parameters (pressure and temperature), and combined with rules in the knowledge graph, reasoning is performed. If insufficient glue layer height is detected and the current pressure is lower than 80% of the standard value, the confidence level of insufficient pressure leading to glue replenishment success rate is queried in the knowledge graph. If it is higher than 80%, it is judged as a secondary defect, and a glue replenishment strategy is generated for re-dispensing and re-judgment. Otherwise, it is judged as a primary defect and rejected. Primary defects represent serious defects that affect product performance, while secondary defects represent minor, repairable defects.

[0127] Based on the multimodal fusion feature vector and real-time dispensing process parameters (pressure and temperature), reasoning is performed using rules in the knowledge graph. Specifically, the proportion of times a rule is hit and processed effectively is used out of the total number of hits. Rules with a confidence level below 70% are marked, triggering a retraining process of historical cases. This optimizes the decision rules in the knowledge graph, improves decision accuracy, and finally outputs the corrected quality level and processing strategy, directly guiding the actions of the execution module.

[0128] Specifically, the adaptive configuration method includes:

[0129] Multiple types of sensors are arranged at the outlet of the glue dispensing area to comprehensively monitor environmental interference parameters. High-precision temperature, humidity, light, and wind speed sensors are used to collect and process environmental parameters in the detection area in real time. For example, according to the layout characteristics of the detection station, the temperature sensor is installed 30cm away from the heat-generating equipment to avoid interference from the equipment's self-heating. The humidity sensor and light sensor are integrated and installed at the front end of the detection head bracket, 50cm away from the workpiece surface. The light sensor's photosensitive surface is vertically upward to accurately collect the ambient light intensity. The wind speed sensor is installed on the top beam of the production line, 1m away from the detection area and horizontally facing the airflow direction to ensure that the airflow speed in the detection area can be captured in real time.

[0130] Each environmental parameter is subjected to a 5-point median filter, and the median of each parameter is calculated to remove impulse noise, generating an environmental parameter vector containing temperature, humidity, light intensity and wind speed, providing real-time environmental data support for subsequent laser power adjustment;

[0131] Obtain the substrate type mark of the finished glued product, and set the base power according to the substrate type. For example, set it to 20mW for high reflective substrates (such as aluminum toothbrush handles) to reduce the influence of specular reflection, set it to 30mW for transparent adhesive layers (such as UV adhesives) to ensure sufficient signal penetration strength, and set it to 25mW for ordinary substrates (such as ABS plastics).

[0132] The environmental parameters are normalized, and the influence of temperature, humidity, light and wind speed on the laser signal is comprehensively considered. The environmental impact factor is calculated by configuring coefficients.

[0133] Based on a pre-defined fuzzy rule base, the environmental impact factor and substrate type are used as inputs. The power adjustment rate is calculated through fuzzy inference. For example, when a high-reflectivity substrate is detected and the environmental impact factor exceeds 0.6, the emission power is reduced by 5%. When a transparent adhesive layer is detected and the humidity exceeds 60%RH, the emission power is increased by 8%. The actual emission power is calculated based on the base power and adjustment rate. At the same time, it is necessary to ensure that the actual emission power is within the pre-defined power limit range. The adjusted laser power value is generated to directly control the output of the laser driver and ensure the stability of the reflected signal quality received by the sensor.

[0134] By using hardware trigger signals and timestamp matching technology, the point cloud stream data of the line laser sensor and the data stream of the environmental sensor are spatiotemporally synchronized. The laser sensor and the environmental sensor are triggered by the same GPIO pulse to ensure that the timestamp error between the two does not exceed ±1μs. At the same time, the industrial camera is connected through a synchronization line to achieve strict alignment between the image frame and the point cloud frame.

[0135] For environmental data with mismatched timestamps, a linear interpolation method is used to generate synchronous environmental parameters corresponding to the timestamps of the laser point cloud, ensuring that each point cloud data corresponds to the real-time environmental state. In the data cleaning stage, invalid points outside the range of the distance sensor, as well as noise points and saturation points with abnormal reflection intensity, are filtered out to generate time-aligned structured point cloud data containing three-dimensional coordinates, reflection intensity, and synchronous environmental parameters. This data serves as the input for subsequent processing modules, providing an accurate raw data foundation for point cloud preprocessing and three-dimensional reconstruction.

[0136] In the detection of substrates with reflectivity exceeding a preset value, by installing orthogonal polarizers at the laser emitter and receiver respectively, 80% of the specular reflection light is filtered out, leaving only the scattered reflection light, which effectively improves the integrity of the contour point cloud.

[0137] For transparent adhesive layers, pulse code modulation technology is used to load a 1kHz pulse sequence onto the laser signal. The receiver calculates the time delay using a cross-correlation function to distinguish the reflected signals from the adhesive layer surface and the substrate.

[0138] By combining multispectral detection technology, a laser that simultaneously satisfies the sensitive wavelength of the emitting adhesive layer and the transmission wavelength of the substrate is used. By separating signals of different wavelengths through a bandpass filter, the problem of signal transmission confusion in the transparent adhesive layer is solved.

[0139] Based on Mie scattering theory, a compensation model for dust concentration and signal attenuation is established to correct the signal attenuation caused by dust scattering, generating an effective reflection signal after anti-interference, providing high-quality input data for subsequent noise suppression and 3D reconstruction, and significantly improving the detection accuracy of complex material surfaces.

[0140] Example 2

[0141] Please see Figure 5 Another embodiment of the present invention provides a method for quality inspection of laser-guided injection contours, comprising the following steps:

[0142] The toothbrush passes through the glue injection area to perform glue injection, generating a glue-injected finished product. A laser sensor emits a laser beam to the glue-injected finished product and receives the laser signal. After median filtering and normalization, the laser power is adjusted in combination with the substrate type and fuzzy rule library. Point cloud and environmental data are synchronized through hardware triggering and timestamp matching technology, and structured point cloud data is generated.

[0143] The point cloud data is converted and invalid points are removed. The gray values ​​are mapped by local normalization. The median filtering and random sampling consistency algorithm are used to remove noise and separate the substrate and adhesive layer edges. The signal attenuation of the transparent adhesive layer is compensated according to the Beer-Lambert law. The surface is reconstructed using the ball-pivot algorithm to generate a parameterized surface model and grayscale image, and a high-precision three-dimensional image is generated.

[0144] The vertex coordinates of the curved surface model are normalized to generate a geometric tensor, the grayscale image is scaled to generate a grayscale tensor, geometric and texture features are extracted and fused, a meta-learning model is used to quickly adapt to new adhesive materials, and an LSTM correlation model is combined to dynamically adjust the qualified threshold and output the initial quality judgment result.

[0145] For products that are initially deemed unqualified, dynamic ROIs are defined based on defect labels, point cloud data of suspicious areas are voxelized, and features are fused to construct a graph structure. Multimodal features are extracted through a graph convolutional network, and the defect type and processing strategy are inferred by combining knowledge graph rules.

[0146] Based on the initial quality assessment and detailed defect assessment results, a differentiated processing strategy is generated, and external equipment is linked to perform calibration, glue replenishment, or rejection operations to achieve closed-loop control of inspection and production.

[0147] In summary, this invention collects point clouds and environmental parameters, adjusts laser power based on substrate type and fuzzy rules, synchronizes data, and cleans to generate structured point clouds. It then denoises, separates the substrate and adhesive layer edges, compensates for signal attenuation in the transparent adhesive layer, reconstructs the surface to generate a 3D model and grayscale image, normalizes to generate geometric and grayscale tensors, extracts fusion features, dynamically adjusts thresholds, and outputs preliminary quality judgment results. For defective products, it delineates suspicious areas, extracts multimodal features through voxelization using multiple techniques, and uses knowledge graphs to infer defect types. Based on the result generation strategy, it links equipment to achieve closed-loop control, improving detection accuracy and automation.

[0148] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A quality inspection system for laser-guided adhesive injection contours, characterized in that, include: Acquire the reflected point cloud signal of the glued product, construct the point cloud data stream, fit the point cloud data, and generate a parametric surface model and grayscale image; The vertex coordinates of the curved surface model are normalized to generate a geometric tensor, the grayscale image is scaled to generate a grayscale tensor, geometric and texture features are extracted and fused to perform cross-layer fusion of features at different scales, and a fused feature vector is generated. A pre-trained lightweight ResNet-18 model is used to generate a probability distribution by fine-tuning the last three fully connected layers for different adhesives and substrates. The quality level is determined based on the probability distribution and dynamic threshold, and the initial quality judgment result is output. For products that are initially deemed unqualified, dynamic ROIs are defined based on defect labels, point cloud data of suspicious areas are voxelized, and features are fused to construct a graph structure. Multimodal features are extracted through a graph convolutional network, and the defect type and processing strategy are inferred by combining knowledge graph rules to generate detailed judgment results. The steps for outputting the initial quality assessment results include: Set initial parameters The MAML algorithm is used to train a meta-learning model, samples of the new adhesive material are collected, fused feature vectors are extracted, and the model parameters are updated according to the meta-learning algorithm to obtain model parameters suitable for the new adhesive material. ; The process parameters of glue injection are collected in real time, and the process parameters and historical glue layer height data are input into the constructed correlation model to output the predicted glue layer height under the current process parameters. Calculate the deviation pressure between the current injection pressure and the standard pressure. If the absolute value of the deviation pressure does not exceed the preset deviation threshold, the process parameters have not fluctuated; otherwise, the process parameters have fluctuated, and the qualified glue height threshold is dynamically adjusted to output the allowable range of glue layer height. Load the adapted model parameters using the student model The fusion feature vector of the glued product is obtained and multi-class probability distributions are output through a fully connected layer; If the predicted probability of the defect category exceeds the preset defect threshold and the adhesive layer height exceeds the allowable range, the product is judged as unqualified; otherwise, it is judged as qualified. The output quality assessment results include the quality grade of the finished glue-filled product and the corresponding preliminary defect labels.

2. The quality inspection system for laser contouring of glue dispensing according to claim 1, characterized in that: The steps for generating parametric surface models and grayscale images include: The initial point cloud data stream is converted into a unified format, invalid points are removed, and a point cloud dataset is formed. In the formula, The number of point clouds to complete a full scan. For the first point cloud dataset One point; For point Select the point Find the points with the smallest Euclidean distance between them to define a local neighborhood; Obtain the maximum and minimum values ​​of the reflection intensity within the local neighborhood, and calculate the value of the point through linear transformation. grayscale value And save it to the cloud dataset mentioned above; Sort the gray values ​​within the local neighborhood in ascending order and calculate the median gray value within the local neighborhood. and standard deviation ; Filter out the points in the point cloud dataset that satisfy the following conditions The points are then removed to form an effective contour point cloud.

3. The quality inspection system for laser contouring of glue dispensing according to claim 2, characterized in that: The steps for generating parametric surface models and grayscale images also include: Set the interior point threshold and the maximum number of iterations, and randomly select three points from the effective contour point cloud to calculate the plane equation and obtain the initial fitting plane; Calculate the distance from points in the effective contour point cloud to the fitting plane, retain points whose distance does not exceed the inlier threshold, and define them as inliers; The calculation is iterated continuously until the maximum number of iterations is reached, thus obtaining the optimal plane and its interior point set. When the transparent adhesive layer is detected, the thickness of the adhesive layer is calculated according to the Beer-Lambert law; Thickness compensation is used to compensate for the intensity of the reflected signal penetrating to the substrate, and the original intensity in the inner point set is replaced with the compensated reflection intensity to generate a pure point cloud. Set the sphere radius, generate a triangular mesh, define the initial control matrix and node vectors, optimize the surface parameters using the least squares method, and generate a parametric surface model. ;in, , For surface parameters; The grayscale values ​​in the prime pure point cloud are mapped to the surface of the triangular mesh through vertex attributes to generate a grayscale image containing surface texture information.

4. The quality inspection system for laser contouring of glue dispensing according to claim 3, characterized in that: The steps for generating the fused feature vector include: The vertex coordinates of the parameterized surface model are normalized to generate a geometric tensor, and the grayscale image is bilinearly interpolated to generate a grayscale tensor. Calculate the position of the vertex in the geometric tensor. axis, axis, The edge tensor is obtained by combining the gradient along the axial direction and the edge intensity. Each point in the geometric tensor is mapped to the corresponding voxel, and the point cloud information in each voxel is statistically analyzed to form voxelized 3D data. Local features and spatial structure information are extracted from the voxelized 3D data to obtain a feature map. Geometric features are then extracted from the feature map, normalized, and a geometric feature vector is formed. The grayscale tensor is processed to generate a multi-scale feature map, and a global average pooling operation is performed to generate a texture feature vector.

5. The quality inspection system for laser contouring of glue dispensing according to claim 4, characterized in that: The steps for generating the fused feature vector also include: For each point in the geometric tensor Calculation points The Euclidean distance between the point and other points in the geometric tensor is selected. The closest Each point forms a neighborhood set; Substitute the coordinates of the points in the neighborhood set into the constructed plane equations, and use least squares plane fitting to obtain an overdetermined linear system of equations. Repeat the neighborhood selection and least squares plane fitting steps described above until all points in the geometric tensor are traversed, resulting in the set of normal vectors of the point cloud, which together form the normal vector tensor. Based on the spatial distribution of point clouds in the parametric surface model, the relationship between adjacent points is defined, the angle between adjacent normal vectors is calculated, and points with an angle greater than a preset threshold are marked as contour points; otherwise, they are marked as non-contour points, and a mask vector is generated. For contour points, let For non-contour points, let , The number of midpoints of the geometric tensor; The weighting coefficients for the contour points are configured as follows: The weighting coefficient for non-contour points is Canny detection is used to identify edge regions of the grayscale image, and the weight coefficients for these edge regions are configured as follows: The geometric feature vector and the texture feature vector are fused using weighting coefficients to generate a fused feature vector.

6. The quality inspection system for laser contouring of glue dispensing according to claim 5, characterized in that: The steps to generate detailed judgment results include: Obtain data on defective products and extract geometric feature thresholds based on the initial defect labels; Based on the geometric feature threshold, suspicious regions corresponding to different defect types are divided, and sub-images corresponding to the suspicious regions are cropped from the grayscale image. The point cloud data of the suspected defect regions and the corresponding grayscale image regions are output. Voxelization and position encoding are performed on the point cloud data of suspected defect areas to generate voxel-level 3D defect segmentation masks. Features are captured using a multi-head self-attention mechanism, and after fusing Transformer and U-Net++ features, defect boundaries are generated using CRF. Using the defect boundary points as nodes, nodes with an Euclidean distance less than a preset threshold are connected, and the cosine of the angle between the node normal vectors is used as the edge weight to construct the defect graph structure. The feature information of nodes and their neighbors is aggregated and calculated to generate graph-level feature vectors, which are then concatenated with the geometric features to generate multimodal fusion feature vectors. Based on the multimodal fusion feature vector and real-time glue injection process parameters, and combined with the rules in the preset knowledge graph, the unqualified products are classified into first-level defects and second-level defects. First-level defects are eliminated, and second-level defects are re-injected. The decision rules of the knowledge graph are periodically optimized to output the correction results.

7. The quality inspection system for laser contouring of adhesive dispensing according to claim 6, characterized in that, Also includes: Read the point cloud data stream, set the adaptive configuration method, obtain environmental parameters, monitor environmental parameters in real time, and dynamically adjust the laser emission power.

8. The quality inspection system for laser contouring of adhesive dispensing according to claim 7, characterized in that: The adaptive configuration method includes: Environmental parameters are collected by sensors placed at the outlet of the dispensing area, and an environmental parameter vector is generated by median filtering. Based on the substrate type, the base power is set, the environmental parameters are normalized, and the impact factor is calculated. The power regulation rate is determined by combining the fuzzy rule base, and the actual transmission power is calculated. Hardware triggering and timestamp matching techniques are used to synchronize point cloud and environmental data, interpolate mismatched data, and generate structured point cloud data. In the detection of substrates with reflectivity exceeding a preset value, a polarizer is installed, and the transparent adhesive layer uses pulse code modulation and multispectral technology. A dust compensation model is established based on Mie scattering theory to correct the signal.

9. A method for quality inspection of laser-cut adhesive contours, used to implement the laser-cut adhesive contour quality inspection system as described in any one of claims 1-8, characterized in that, include: A laser sensor is used to emit a laser to the finished glue-filled product and receive the laser signal. Environmental parameters are monitored in real time, the laser emission power is dynamically adjusted, and structured point cloud data is generated. The point cloud data is converted into a new format and invalid points are removed. The point cloud data is then fitted to generate a parametric surface model and a grayscale image. The extracted geometric features and texture features are fused together, a meta-learning model is used to adapt to the new adhesive material, and the qualified threshold is dynamically adjusted by combining the correlation model to output the initial quality judgment result. For finished glued products that initially fail the inspection, a defect assessment is conducted to generate detailed assessment results.

Images