A deep learning-based powder sub-packaging quality visual detection method

Abstract

This invention relates to the field of deep learning technology, specifically to a deep learning-based visual inspection method for powder packaging quality. The method includes simultaneously acquiring visible light transmission images and surface depth point clouds of the sealing area after heat sealing; constructing an anisotropic differential operator along the airflow direction to extract the solid-gas composite inclusion profile of tungsten carbide powder and determine the composite positioning points; performing elevation attenuation gradient analysis along the heat sealing embossing direction; adaptively cutting a local depth matrix using an eccentric ellipse as the boundary; calculating the film protrusion height, negative elevation region mask, and negative elevation integral through thin-plate spline fitting to a reference surface; and combining these into a local tent deformation feature matrix. Finally, one route uses an attenuation neural network to output a sealing risk index, while another route uses a tear risk prediction neural network to output a dynamic tear risk score. Threshold comparison is then used to determine whether the sealing performance and strength of the seal meet the standards. This invention achieves powder sealing quality assessment through multimodal visual fusion.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, specifically to a deep learning-based visual inspection method for powder packaging quality. Background Technology

[0002] Computer vision and deep learning technologies have been widely applied in industrial product quality inspection. Taking the encapsulation structure formed by hot-pressing multilayer flexible films as an example, the inspection process typically involves acquiring surface images through optical sensors, and then using methods such as edge detection operators, threshold segmentation, or convolutional neural networks to extract two-dimensional texture and shape features, ultimately providing a classification result.

[0003] However, these types of objects may contain tiny hard foreign objects inside, and their surfaces may have embossing. It is difficult to obtain three-dimensional elevation information reflecting the spatial undulation of the surface using only a single-modal two-dimensional visible light image. In other words, the surface deformation data carried by two-dimensional images is insufficient to support an accurate judgment of packaging quality.

[0004] Another issue lies in data fusion. Conventional deep learning detection models mostly accept fixed-size regular grid image inputs to complete object detection or semantic segmentation tasks. When it is necessary to simultaneously utilize image data from multiple modalities, such as 2D geometric contours and 3D depth point clouds, and perform joint inference with external operating condition constants (such as scalar parameters like pressure and temperature), the shortcomings of existing processes become apparent: On the one hand, image cropping usually uses fixed windows or whole-image inputs, lacking a mechanism for adaptive cropping based on the anisotropic physical structural characteristics of the detected object, easily introducing a large amount of useless background information; on the other hand, existing feature fusion networks, when processing cross-modal visual feature tensors and non-visual scalar constants, do not establish a corresponding relationship between the physical properties of the detected object, making it difficult for the model to output assessment indicators that can quantitatively reflect the sealing risk or stress-induced damage risk under stress conditions, even if the model extracts visual features.

[0005] Therefore, how to effectively map specific physical performance evaluation parameters of the detected object from visual shape data using deep learning networks is a technical problem that needs to be solved.

[0006] To address this, a deep learning-based visual inspection method for powder packaging quality is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a deep learning-based visual inspection method for powder packaging quality. Through multimodal visual fusion and deep neural networks, it achieves powder sealing quality assessment. The method includes first acquiring visible light transmission images and surface depth point clouds of the heat-sealing area; then extracting the solid-gas composite inclusion profile of tungsten carbide powder and determining the composite positioning points using an anisotropic differential operator constructed along the airflow direction; next, performing elevation attenuation gradient analysis along the direction of the heat-sealing embossing; and adaptively extracting the local depth matrix using an eccentric elliptical boundary. After fitting a reference surface with a thin-plate spline, the film protrusion height, negative elevation region mask, and negative elevation integral are calculated and assembled into a local tent deformation feature matrix. Finally, the process is divided into two paths: one uses an attenuation neural network to predict the vacuum attenuation rate, and the other uses a prediction neural network to output a dynamic tear risk score, comprehensively determining whether the sealing performance and strength of the heat seal meet the standards.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based visual inspection method for powder packaging quality includes: Receive the heat sealing completion signal and acquire visible light transmission image and surface depth point cloud image of the heat-sealed area; The visible light transmission image is input into the anisotropic differential operator to extract particle contour data and composite positioning point data; the composite positioning point data is mapped to the surface depth point cloud image, and local depth matrix data is extracted according to preset cropping parameters. The local depth matrix data is input into the surface fitting algorithm to calculate the maximum normal height difference. The local depth matrix is ​​then subtracted element by element from the fitting reference surface to obtain the local relative elevation matrix. From the local relative elevation matrix, the film protrusion height data, the mask data of the negative elevation region, and the negative elevation integral are extracted. The particle contour data, the film protrusion height data, and the negative elevation integral are then concatenated to generate the local tent deformation feature matrix data. A connected region search is performed on the mask data of the negative elevation region along the sealing width direction to extract the continuous path through the heat seal and calculate the leakage channel vector data across the seal. The data is then input into the attenuation neural network and outputs the predicted vacuum attenuation rate data. The equivalent tensile stress constant data of the target lifting load and the elevation gradient along the force direction of the film protrusion height data are fused to generate load-morphology gradient proxy matrix data. This data is then input into the prediction neural network and outputs the dynamic tear risk score data. The predicted vacuum decay rate data and dynamic tear risk score data are compared with preset thresholds to output the quality judgment label data of the heat-sealed area.

[0009] Preferably, the process of receiving the heat-sealing completion signal and acquiring the visible light transmission image and surface depth point cloud image of the heat-sealed area includes: receiving a heat-sealing completion trigger signal sent by the main control device; responding to the heat-sealing completion trigger signal, controlling the transmission light source located below the heat-sealed area to emit light, and controlling the first camera located directly above the heat-sealed area to acquire the light penetrating the heat-sealed area, generating the visible light transmission image data; synchronously controlling the structured light generator located above the heat-sealed area to project a structured light pattern onto the heat-sealed area, controlling the second camera arranged at a preset relative angle to the structured light generator to acquire the image of the heat-sealed area projected with the structured light pattern, and performing three-dimensional reconstruction based on the acquired image to generate the surface depth point cloud image data.

[0010] Preferably, the process of generating particle contour data and composite positioning point data includes: inputting the visible light transmission image into an anisotropic differential operator based on the direction of the airflow line, extracting the absolute extinction region formed by the tungsten carbide hard core, and the annular micro-cavity region formed by the micro-airflow blockage around the absolute extinction region, generating solid-gas composite inclusion contour data as the particle contour data; calculating the convex hull polygon of the absolute extinction region, extracting the coordinates of the sharp vertex of the polygon with the largest convex distance along the heat-sealing stretching direction; and vector superimposing the coordinates of the sharp vertex of the polygon with the geometric center coordinates of the annular micro-cavity region to generate the composite positioning point data.

[0011] Preferably, the process of extracting local depth matrix data according to preset truncation parameters includes: mapping the composite positioning point data to the surface depth point cloud image according to camera calibration parameters to locate the three-dimensional coordinates of the deformation vertex; extracting the heat-sealing embossing direction vector within a preset range around the three-dimensional coordinates of the deformation vertex; extracting a first elevation attenuation gradient along the direction of the heat-sealing embossing direction vector, and extracting a second elevation attenuation gradient along the direction perpendicular to the direction of the heat-sealing embossing direction vector; using the positions where the first elevation attenuation gradient suddenly becomes zero and the positions where the second elevation attenuation gradient suddenly becomes zero as boundary points to construct eccentric elliptical boundary data; and truncating local point clouds from the surface depth point cloud image along the eccentric elliptical boundary data to generate the local depth matrix data.

[0012] Preferably, the process of generating film protrusion height data, negative elevation region mask data, and negative elevation integral includes: inputting local depth matrix data into a least-squares thin-plate spline fitting algorithm to construct heat-sealing reference surface data that fits the undeformed base surface of the heat-sealing film; performing matrix subtraction between the local depth matrix data and the heat-sealing reference surface data to generate local relative height difference matrix data; extracting the maximum positive value from the local relative height difference matrix data to generate film protrusion height data; extracting the set of grid coordinates and their relative elevation values ​​of negative elevation regions located around the film protrusion height data and with values ​​less than zero from the local relative height difference matrix data to generate negative elevation region mask data; performing integral calculation along the normal of the heat-sealing reference surface on the negative elevation region, and using the cumulative value of the product of the absolute value of the relative elevation within the negative elevation region and the physical area of ​​the grid as the negative elevation integral.

[0013] Preferably, the process of generating local tent deformation feature matrix data includes: extracting geometric feature vectors from particle contour data and normalizing the geometric feature vectors to generate normalized particle morphology data; mapping the film protrusion height data and the negative elevation integral according to a preset ratio to generate normalized spatial deformation data; and splicing the normalized particle morphology data and the normalized spatial deformation data in a tensor dimension according to a preset channel order, and aligning the spatial center based on the composite positioning point data to generate local tent deformation feature matrix data.

[0014] Preferably, the process of outputting the predicted vacuum attenuation rate data includes: establishing a neighborhood graph structure for the negative elevation region mask data with a grid as the basic unit along the width direction from the inner to the outer side of the heat seal; performing a connected component search to extract the continuous recessed path that runs through the heat seal; calculating the channel morphology feature vector along the extension direction of the continuous recessed path to generate the cross-seal leakage channel vector data; inputting the cross-seal leakage channel vector data into the attenuation neural network for mapping, calculating and outputting the predicted vacuum attenuation rate data.

[0015] Preferably, the process of outputting dynamic tear risk score data includes: extracting the local relative elevation matrix corresponding to the film protrusion height data; calculating the partial derivative of the local relative elevation matrix in the lifting force direction to generate force-direction elevation gradient matrix data; multiplying the equivalent tensile stress constant data of the target lifting load with the force-direction elevation gradient matrix data at corresponding positions to generate load-morphology gradient proxy matrix data; and inputting the load-morphology gradient proxy matrix data into a prediction neural network to calculate and output the dynamic tear risk score data.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This method incorporates the absolute extinction characteristics caused by tungsten carbide powder and the annular micro-cavities formed by airflow obstruction into the particle profile description framework, rather than focusing solely on the two-dimensional geometry of the particle itself. Based on this, the sharpest vertex along the stretching direction is identified from the convex hull polygon and vector-superimposed with the geometric center of the air cavity to locate the centroid. The advantage of this approach is that the generated local centroid data simultaneously carries two layers of information: the physical morphology of the particle and the location information of the microscopic air channels surrounding the particle. This is equivalent to upgrading traditional planar geometric positioning to a joint extraction of the solid-gas composite structure. The positioning point is no longer just a geometric center, but a composite anchor point that can simultaneously point to high-risk areas of puncture and tearing and weak sealing areas.

[0017] 2. When extracting the local depth matrix, instead of simply using a fixed-size window, we first extract the heat-sealing embossing direction vector of the region surrounding the deformation vertex. Then, we calculate the elevation attenuation gradient along this vector direction and its perpendicular direction, finding the point where the gradient converges to zero as the boundary point. We then use these four boundary points to fit an asymmetric, eccentric ellipse. The underlying logic of this approach is to utilize the anisotropy of deformation propagation of the heat-sealed film under the constraint of the textured structure. Deformation extends further in the embossing direction and attenuates faster in the perpendicular direction, naturally conforming to this physical law in the major and minor axes of the ellipse. The final extracted local depth matrix not only encompasses the actual deformation range caused by the hard particles supporting the film and preserves effective microscopic morphological features, but also filters out background data from flat, deformation-free areas, reducing unnecessary computation in subsequent processing.

[0018] 3. By calculating the partial derivative of the local relative elevation matrix corresponding to the film protrusion height data in the direction of lifting force, a force-direction elevation gradient matrix representing the steepness of the film is generated. This matrix is ​​then multiplied by the equivalent tensile stress constant of the target lifting load according to position. The resulting load-morphology gradient proxy matrix can combine the load strength with the morphological abrupt change in the direction of force, providing proxy features for the subsequent risk prediction model that reflect the relative risk ranking of each local area under the stress condition. Attached Figure Description

[0019] Figure 1 A schematic diagram of a deep learning-based visual inspection method for powder packaging quality provided by the present invention; Figure 2 This is a schematic diagram of the process for generating particle contour data and composite positioning point data provided by the present invention; Figure 3 This is a schematic diagram of the process for extracting local depth matrix data according to preset truncation parameters provided by the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figures 1 to 3 This invention provides a deep learning-based visual inspection method for powder packaging quality, the technical solution of which is as follows: A deep learning-based visual inspection method for powder packaging quality includes: Receive the heat sealing completion signal and acquire visible light transmission image and surface depth point cloud image of the heat-sealed area; The visible light transmission image is input into the anisotropic differential operator to extract particle contour data and composite positioning point data; the composite positioning point data is mapped to the surface depth point cloud image, and local depth matrix data is extracted according to preset cropping parameters. The local depth matrix data is input into the surface fitting algorithm to calculate the maximum normal height difference. The local depth matrix is ​​then subtracted element by element from the fitting reference surface to obtain the local relative elevation matrix. From the local relative elevation matrix, the film protrusion height data, the mask data of the negative elevation region, and the negative elevation integral are extracted. The particle contour data, the film protrusion height data, and the negative elevation integral are then concatenated to generate the local tent deformation feature matrix data. A connected region search is performed on the mask data of the negative elevation region along the sealing width direction to extract the continuous path through the heat seal and calculate the leakage channel vector data across the seal. The data is then input into the attenuation neural network and outputs the predicted vacuum attenuation rate data. The equivalent tensile stress constant data of the target lifting load and the elevation gradient along the force direction of the film protrusion height data are fused to generate load-morphology gradient proxy matrix data. This data is then input into the prediction neural network and outputs the dynamic tear risk score data. The predicted vacuum decay rate data and dynamic tear risk score data are compared with preset thresholds to output the quality judgment label data of the heat-sealed area.

[0022] Example 1: Furthermore, the process of receiving the heat-sealing completion signal and acquiring the visible light transmission image and surface depth point cloud image of the heat-sealed area includes: receiving a heat-sealing completion trigger signal sent by the main control device; responding to the heat-sealing completion trigger signal, controlling the transmission light source located below the heat-sealed area to emit light, and controlling the first camera located directly above the heat-sealed area to acquire the light penetrating the heat-sealed area, generating the visible light transmission image data; synchronously controlling the structured light generator located above the heat-sealed area to project a structured light pattern onto the heat-sealed area, controlling the second camera arranged at a preset relative angle to the structured light generator to acquire the image of the heat-sealed area projected with the structured light pattern, and performing three-dimensional reconstruction based on the acquired image to generate the surface depth point cloud image data.

[0023] During system operation, when the packaging bag heat-sealing mechanism completes the pressing action and resets, the main control equipment of the production line sends a hard trigger level signal to the synchronous controller of the vision inspection system through the hardware I / O interface as a heat-sealing completion trigger signal. After receiving the signal, the synchronous controller generates two hardware trigger pulses with a time interval, which sequentially activate the transmission imaging optical path and the structured light imaging optical path, respectively. The time interval between the two pulses is configured to be greater than the sum of the preset exposure time of the camera and the sensor data readout time in the first transmission imaging optical path, so as to avoid spectral crosstalk between the two light sources in the same physical space through the time-division stroboscopic acquisition method.

[0024] For the transmission imaging process, the synchronous controller outputs the first pulse to illuminate the high-brightness LED array backlight installed directly below the heat-sealing area of ​​the packaging bag, and simultaneously triggers the first array camera with its optical axis vertically downward and aligned with the heat-sealing opening. The uniform light beam emitted by the backlight penetrates the semi-transparent flexible heat-sealing film from bottom to top. The opaque tungsten carbide powder inside the film physically blocks the light beam. The first array camera captures the light that penetrates the film within a fixed exposure time and converts it into a two-dimensional grayscale or RGB pixel matrix to generate visible light transmission image data. It should be noted that this method is applicable to detection scenarios where tungsten carbide powder exists in the heat-sealing area as discrete micro-particles, rather than as a continuous powder layer covering the entire sealing area. Specifically, the diameter of a single tungsten carbide foreign particle does not exceed one-tenth of the width of the heat-sealing opening, and the cumulative occlusion area of ​​all foreign particles in the transmission image does not exceed 20% of the total image area.

[0025] After this processing, most of the sealed area still retains effective transmission grayscale information. In this imaging scenario, the location of the tungsten carbide particles appears as an absolute extinction patch with a grayscale value close to zero in the transmission image, while light can penetrate the unblocked heat-sealed film area between the particles, resulting in a significantly higher grayscale value. This creates a high-contrast grayscale hierarchy, providing sufficient image information for subsequent particle contour extraction.

[0026] Upon entering the surface depth acquisition process, the synchronous controller emits a second pulse after the set time interval, triggering a structured light generator installed above the heat-sealing area. This generator, based on a digital light processing projection module, continuously projects multiple frames of sinusoidal grating fringe patterns onto the heat-sealing surface. These fringes contain different frequencies, each group undergoing a four-step phase shift. Due to deformation and the presence of protrusions on the heat-sealing surface, the projected grating fringes are physically modulated—the originally parallel fringes exhibit geometric deformation and phase shift at these locations. Simultaneously with the structured light projection, a second camera is synchronously triggered for continuous acquisition. This camera maintains a preset baseline angle of 30 to 45 degrees with the structured light generator. The second camera continuously acquires multiple frames of fringe image sequences modulated by surface deformation. First, a four-step phase shift algorithm extracts the encapsulated phase of each pixel from each group of fringes. Then, a multi-frequency heterodyne phase unpacking algorithm fuses the phase information of fringes from different frequencies to calculate the absolute phase distribution of each pixel within the target area.

[0027] In the 3D reconstruction stage, the structured light generator and the second camera are constructed as a camera-projector dual-node geometric model. The camera intrinsic parameter matrix, the projector intrinsic parameter matrix, and the extrinsic parameter rotation and translation matrix between them are called in advance through the calibration board offline calibration. Based on the epipolar geometric constraints and the triangulation model, the two-dimensional image coordinates of each pixel along with its absolute phase value are mapped to the set of three-dimensional physical coordinate points X, Y, and Z in the spatial rectangular coordinate system, thereby obtaining the surface depth point cloud image data.

[0028] Heat-sealing films possess a certain optical penetration depth in the visible light band (400-700nm). After incident structured light fringes enter the film, some scattered light escapes from the vicinity of the projection point, leading to a decrease in the fringes contrast received by the camera. This method addresses this issue on two levels. The first level involves the selection of the light source wavelength: using a shorter wavelength structured light projection source, such as the blue light band (440-460nm) or the near-ultraviolet band (385-405nm). In these bands, the optical absorption coefficient of the heat-sealing film material increases, the penetration depth decreases accordingly, and the subsurface scattering component is effectively suppressed. The second level involves the phase resolution strategy: using the multi-frequency heterodyne method to demodulate the high-frequency fringe components. High-frequency fringes are inherently more robust to contrast attenuation caused by scattering. Under blue light projection conditions, the modulation transfer function of the high-frequency fringes (e.g., 64 periods / field of view) after reflection from the film surface is pre-calibrated and confirmed to meet the basic threshold for fringe contrast in the four-step phase-shift algorithm.

[0029] By setting the baseline angle between the structured light generator and the second camera within the range of 30 to 45 degrees, the camera's optical axis naturally avoids the main lobe direction of specular reflection. In most imaging areas, the contribution of the specular reflection component to the total received light intensity can be controlled to below 20%. For a few pixels that still exhibit high light saturation in certain areas, they are marked as invalid regions during the phase calculation stage and then interpolated from neighboring valid pixels to complete the region.

[0030] Regarding the imaging impact of tungsten carbide powder particles embedded in the thin film region, it needs to be considered in conjunction with the aforementioned transmission imaging scenario. The cumulative area of ​​the powder particles does not exceed 20% of the total image area, and the powder particles themselves exhibit diffuse reflectivity, displaying local stripe modulation characteristics different from the surrounding thin film region in the structured light stripe reflection image, without causing the loss of global stripe information. A further comparison in terms of size: the typical thickness of heat-sealed films is between 50μm and 200μm, and the protrusion height (i.e., deformation peak) formed by the particles supporting the film in the detection scenario is typically 2 to 5 times the film thickness, or on the order of 100μm to 1mm. At a working distance of 300mm, the measurement uncertainty in the depth direction of structured light 3D reconstruction has been calibrated and determined; the size of the protrusion is more than an order of magnitude larger than the measurement error, and the current measurement accuracy meets the detection requirements.

[0031] Next, we proceed to the feature extraction stage. The acquired visible light transmission image is input into an anisotropic differential operator based on the direction of the airflow path. This operator simultaneously extracts two types of regions from the image: one is the absolutely extinct region formed by the tungsten carbide hard core, and the other is the annular micro-cavity region surrounding the absolutely extinct region due to microscopic airflow obstruction. The contours of these two types of regions are topologically merged to generate solid-gas composite inclusion contour data, denoted as particle contour data. Subsequently, a convex hull polygon is calculated for the absolutely extinct region. Among all vertices of the convex hull, the vertex with the largest convex distance along the heat-sealing stretching direction is found, and its coordinates are marked as the coordinates of the polygon's sharp vertex. Then, the coordinates of the polygon's sharp vertex are weighted and superimposed with the geometric center coordinates of the annular micro-cavity region to obtain composite positioning point data.

[0032] Further, the process of generating particle contour data and composite positioning point data includes: inputting the visible light transmission image into an anisotropic differential operator based on the direction of the airflow line, extracting the absolute extinction region formed by the tungsten carbide hard core, and the annular micro-cavity region formed by the micro-airflow blockage around the absolute extinction region, generating solid-gas composite inclusion contour data as the particle contour data; calculating the convex hull polygon of the absolute extinction region, extracting the coordinates of the sharp vertex of the polygon with the largest convex distance along the heat-sealing stretching direction; and vector superimposing the coordinates of the sharp vertex of the polygon with the geometric center coordinates of the annular micro-cavity region to generate the composite positioning point data.

[0033] The acquired visible light transmission image data is input into the edge computing nodes deployed in the industrial field for real-time processing. Within the edge computing nodes, the relative position and air pressure parameters of the exhaust nozzle set by the packaging process are first analyzed to establish a two-dimensional airflow direction matrix. Specifically, the edge computing nodes extract the two-dimensional center coordinates of the exhaust nozzle in the image coordinate system and equate it to the potential flow sink in an ideal fluid field. As an engineering approximation, the exhaust nozzle is modeled as the potential flow sink of an incompressible ideal fluid. It should be noted that the actual vacuum pumping process is a high-speed, compressible transient flow in the initial stage, with strong unsteady characteristics in the airflow direction. As the pumping progresses into the middle and later stages, and the pressure inside the packaging bag drops to near the target vacuum level, the velocity of the residual gas decreases, and the Mach number drops below 0.3. At this point, the gas can be approximated as an incompressible fluid, and the flow within the narrow gap of the heat-sealed opening tends towards a laminar stable state. The deviation between the airflow direction calculated by the ideal potential flow model and the actual flow direction is within an acceptable range in engineering. Therefore, the airflow line direction vector angle calculated using the potential flow sink model is suitable for the quasi-steady-state flow condition in the middle and later stages of pumping, serving as a directional reference for the subsequent anisotropic differential operator. Based on the spatial connection vector from each pixel coordinate to the potential flow sink, the pumping airflow line direction vector angle corresponding to each pixel coordinate is calculated. Subsequently, a parameterized anisotropic Gaussian differential operator matrix is ​​constructed with the direction of the airflow vector angle as the main derivative direction.

[0034] It should be noted that the role of the airflow direction vector angle in this scheme is to provide spatial directionality priors for the anisotropic differential operator, rather than for calculating precise flow field physical quantities. The directional selectivity of the anisotropic differential operator has a certain tolerance for deviations in the input angle. The attenuation of the convolution response amplitude caused by directional deviation has been calibrated and confirmed not to substantially affect the extraction quality of the particle-air cavity composite profile. Therefore, the impact of the local directional deviation introduced by the potential flow model on the overall detection performance is limited. Its construction process is as follows: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] A coordinate rotation matrix is ​​introduced to perform a spatial rotation transformation on the XY-axis coordinates of the standard two-dimensional Gaussian distribution function, generating a Gaussian kernel function with its major axis parallel to the airflow direction and its minor axis perpendicular to the airflow direction. A two-dimensional spatial convolution calculation is then performed between the visible light transmission image and this anisotropic differential operator. Since the tungsten carbide hard core has extremely high density and is completely opaque, the absolute extinction region formed by the tungsten carbide hard core is extracted by extracting a set of pixels from the convolution response image whose gradient magnitude is greater than a set edge threshold and whose internal connected component grayscale values ​​are lower than a preset hard occlusion threshold. Then, the boundary of the extracted absolute extinction region is used as the initial... Seed points are used to grow and search outwards along a direction perpendicular to the local airflow line. In the transmission imaging model, the micro-cavities formed by the rigid core blocking the airflow have micron-level characteristic dimensions, which are close to or slightly larger than the wavelength of visible light. According to the Mie scattering theory, when the transmitted light beam passes through the thin film-cavity interface, the refractive index difference between the materials on both sides of the interface (the refractive index of the heat-sealed film is about 1.5, and the refractive index of the air inside the cavity is about 1.0, with a refractive index difference of about 0.5) causes the incident light to be scattered at the boundary of the cavity. The scattered light deviates from the original transmission direction, which manifests as a resolvable attenuation of the local transmitted light intensity at the corresponding pixel position on the camera's focal plane. In an 8-bit grayscale image, the grayscale value of the micro-cavity region is noticeably lower than that of the surrounding non-cavity film region, forming a visible annular grayscale transition band. The edge computing node extracts the continuous pixel band surrounding the absolute extinction region and whose grayscale value is within the preset air cavity refraction threshold range to extract the annular micro-cavity region formed by the micro airflow blockage. Then, a topological union operation is performed on the outer edge point set of the absolute extinction region and the outer boundary point set of the annular micro-cavity region to extract the outermost closed contour chain code after merging, generating solid-gas composite inclusion contour data, which is then stored as the particle contour data.

[0035] For the extracted absolute extinction region, the Graham scan algorithm is used to traverse the edge contour point set of the region, calculate and generate the minimum convex hull polygon surrounding the extinction region, and output the vertex sequence of the polygon. Simultaneously, the unit vector of the heat-sealing stretching direction under the current process setting is acquired, and the coordinates of all vertices in the convex hull polygon vertex sequence are projected onto the spatial line containing the heat-sealing stretching direction unit vector using a dot product. The projected distance value of each vertex in the stretching direction is calculated, and the vertex with the largest projected distance value is extracted using bubble sort, and its coordinates are recorded as the coordinates of the polygon's sharp vertex. .

[0036] The arithmetic mean of the two-dimensional coordinates of all pixels within the annular micro-cavity region is calculated to generate the geometric center coordinates of the annular micro-cavity region. The coordinates of the sharp vertices of the polygon The geometric center coordinates of the annular micro-cavity region Both are converted into position vectors referencing the same origin. In the edge computing nodes, the two position vectors are linearly summed with weights according to the following preset weight coefficients: in, Let be the position vector of the sharp vertices of the polygon. The position vector represents the geometric center coordinates of the annular microcavity region. The preset weighting coefficient ranges from 0 to 1. This weighting coefficient is preset by the process engineer according to the following principle: if the production line is more concerned about the risk of hard foreign object puncture than the risk of micro-cavity leakage (i.e., more concerned about the accurate location of the tear source), then... Set a value greater than 0.5 to offset the centroid position towards the sharp vertex; if the production line prioritizes the risk of leaks due to sealing issues over the risk of punctures, then... Setting it to a value less than 0.5 will shift the centroid position towards the center of the air cavity; the default value is [value missing]. That is, the sharp apex and the center of the air cavity are superimposed with equal weight.

[0037] The composite positioning point data is generated by using the new two-dimensional spatial position vector output after vector superposition based on the above formula.

[0038] Further, the process of extracting local depth matrix data according to preset truncation parameters includes: mapping the composite positioning point data to the surface depth point cloud image according to camera calibration parameters to locate the three-dimensional coordinates of the deformation vertex; extracting the heat-sealing embossing direction vector within a preset range around the three-dimensional coordinates of the deformation vertex; extracting a first elevation attenuation gradient along the direction of the heat-sealing embossing direction vector, and extracting a second elevation attenuation gradient along the direction perpendicular to the direction of the heat-sealing embossing direction vector; using the positions where the first elevation attenuation gradient abruptly becomes zero and the positions where the second elevation attenuation gradient abruptly becomes zero as boundary points to construct eccentric elliptical boundary data; and truncating local point clouds from the surface depth point cloud image along the eccentric elliptical boundary data to generate the local depth matrix data.

[0039] Pre-calibration data between the transmission camera and the structured light camera is acquired. This calibration data includes the intrinsic parameter matrix of the transmission camera and the joint extrinsic translation and rotation matrix between the two camera coordinate systems. Local two-dimensional centroid coordinate data is used as input in the transmission image coordinate system. First, it is multiplied by the inverse of the intrinsic parameter matrix of the transmission camera to convert it into a spatial direction vector in the camera coordinate system. Then, this spatial direction vector is mapped to a three-dimensional ray in the surface depth point cloud image space by combining the joint extrinsic translation and rotation matrix with epipolar geometric constraints. Simultaneously, the RANSAC algorithm is used to perform plane fitting on the globally flat regions in the surface depth point cloud image, constructing the equation as follows: Using the reference surface of the point cloud, in the surface depth point cloud image, with the intersection point of the ray and the reference surface in three-dimensional space as the center, a data traversal search is performed within a cylindrical space of a preset radius. The elevation values ​​of all point cloud data within this cylindrical space are extracted and compared using a bubble sort method. The point cloud coordinates with the largest elevation value are extracted and output as the three-dimensional coordinates of the deformation vertex. .

[0040] A three-dimensional spherical neighborhood with a preset radius is established, centered on the three-dimensional coordinates of the deformed vertex, and a local point cloud within this neighborhood is extracted. A two-dimensional polynomial surface fitting is performed on the elevation distribution of the local point cloud, and the partial elevation derivatives along the X and Y axes of each point within this neighborhood are calculated. and And construct the structure tensor matrix based on the product of partial derivatives. Its mathematical expression is: The structural tensor matrix is ​​subjected to eigenvalue decomposition to extract the eigenvector corresponding to the smallest eigenvalue. This eigenvector is then vertically projected onto the horizontal reference plane of the point cloud to obtain a two-dimensional spatial vector reflecting the main direction of the periodic arrangement of the heat-sealing embossing, which is used as the heat-sealing embossing direction vector.

[0041] Using the horizontal projection point of the three-dimensional coordinates of the deformation vertex as the geometric origin, the direction of the heat-sealing embossing vector is set as the U-axis, and the direction perpendicular to this vector is set as the V-axis, establishing a locally orthogonal coordinate system. Starting from this geometric origin, the elevation values ​​of spatial points are extracted at equal intervals along the positive and negative directions of the U-axis in the point cloud. The curve of elevation value change with the U-axis coordinate is calculated using first-order difference to extract the first elevation attenuation gradient sequence along the direction of the heat-sealing embossing vector. Similarly, the elevation values ​​of spatial points are extracted at equal intervals along the positive and negative directions of the V-axis and first-order difference is performed to extract the second elevation attenuation gradient sequence along the direction perpendicular to the direction of the heat-sealing embossing vector.

[0042] One-dimensional Gaussian smoothing filtering is applied to the first and second elevation attenuation gradient sequences. The filtered gradient sequences are traversed outwards along the U-axis and V-axis from the geometric origin, respectively, to detect the spatial coordinates where the absolute value of the gradient first decreases and remains within a preset near-zero tolerance range. Two near-zero coordinates in the positive and negative directions of the U-axis are extracted as the endpoints of the major axis. and Extract two near-zero coordinate points in the positive and negative directions of the V-axis as the endpoints of the minor axis. and Based on the extracted coordinates of the four endpoints, an asymmetric elliptic equation allowing for center point offset is constructed: Substitute the coordinates of the four endpoints into the equation and solve the system of equations to find the coordinates of the ellipse's center. , and major and minor axis parameters and A closed elliptical trace containing center offset parameters and major and minor axis parameters is constructed by fitting; the ellipse is axis-aligned in the local UV coordinate system, and its direction in the global point cloud coordinate system is determined by the angle between the U axis and the global coordinate axis, generating eccentric elliptical boundary data.

[0043] The eccentric elliptical boundary data is projected onto the two-dimensional reference plane where the surface depth point cloud image is located, forming a closed vertical cylindrical filter boundary. The entire surface depth point cloud image is traversed, and it is determined whether the horizontal projection of each three-dimensional point cloud coordinate lies inside the closed cylindrical filter boundary. Point cloud data outside the boundary is removed, and the internal point cloud set is retained. Two-dimensional gridded resampling interpolation calculation is performed on the retained internal point cloud set according to a preset physical sampling interval, converting the non-uniform discrete point cloud data into a row-column aligned elevation numerical array, generating the local depth matrix data.

[0044] The negative elevation region is integrated along the normal of the heat-sealing reference surface. The negative elevation integral is defined as: the absolute value of the relative elevation at each grid point within the negative elevation region, multiplied by the physical area of ​​that grid, and then the results of all grids are summed.

[0045] Next, background control points are extracted. Edge pixels of a preset width (usually 3 to 5 pixel layers) are taken from the outermost edge of the local depth matrix, and their coordinates and elevation values ​​are recorded as a set of candidate background control points that have not undergone physical deformation. However, before directly selecting these edge pixels, an elevation value consistency check needs to be performed: the standard deviation of the elevation values ​​of the outer pixels is calculated. If the standard deviation exceeds a preset threshold (this threshold is determined after calibration based on the surface roughness of the heat-sealing film material), it indicates that the edge region may be affected by the deformation of adjacent particles. At this time, the outer width is reduced layer by layer inward, and the check is repeated until the standard deviation drops below the threshold, or the outer width is reduced to at least 2 pixel layers. Furthermore, to minimize the impact of adjacent particle deformation on the reference surface fitting, regions at least 3 times the equivalent diameter of the particle at the current particle centroid are preferentially selected as candidate regions for background control points. The above-selected set of background control points is used as input to construct a least-squares thin-plate spline fitting model, with the basic radial basis function taken as (…). ),in Let be the Euclidean distance between the two points, and in the algorithm calculation, it is defined that... Establish the surface interpolation equation consisting of an affine transformation part and a nonlinear deformation part: in, Let be the plane coordinates of the point to be determined. Let be the position vector of the point to be determined. Let i be the position vector of the i-th background control point. This represents the total number of background control points. For affine parameters, These are the weighting coefficients corresponding to each control point.

[0046] In the solution process, the minimum bending energy of the thin plate is used as the constraint condition, and a system of linear equations is constructed in combination with smooth boundary conditions. The weight coefficients corresponding to each control point are solved by matrix operations. Then, the two-dimensional planar grid coordinates of all pixels in the local depth matrix data are substituted into the spline interpolation equation after the solution is obtained, and the reference elevation value of each coordinate point in the local area under the ideal undeformed state is calculated, thereby constructing the heat-sealing reference surface data with the same row and column dimensions as the original matrix.

[0047] The original local depth matrix data With the heat-sealing reference surface data Element-by-element matrix subtraction is performed to generate a local relative height difference matrix containing positive and negative fluctuation information. All numerical elements in this local relative height difference matrix are traversed, and the largest positive value is found through sorting and comparison. This largest positive value is taken as the highest geometric peak value formed by the rigid particle-supported film, directly assigned, and the film protrusion height data is generated. Simultaneously, the two-dimensional grid index coordinates of the largest positive value are recorded.

[0048] Centered on the grid index coordinates where the maximum positive value is located, a circular search neighborhood is defined according to a preset physical diffusion radius. Within this circular search neighborhood, the local relative height difference matrix data is traversed, and the set of grid coordinates with values ​​strictly less than zero is extracted. This set is then designated as the negative height region that cannot be properly aligned due to stretching deformation. Obtain the physical spacing of the gridded resampling set in the local depth matrix generation step, i.e., the physical length of a single grid cell. With physical width In the negative height region Within this model, a discrete summation method is used to calculate the product of the absolute value of the relative height difference and the physical area of ​​a single mesh. A double integral approximation is then applied to the missing volume of the reference surface normal. The discrete integral formula is as follows: in, This is the volumetric scalar value of the negative elevation integral; This is the set of grid coordinates corresponding to the negative height region that cannot be fitted due to stretching deformation; For the region The horizontal and vertical index coordinates of the two-dimensional grid within; This represents the local relative height difference at the corresponding coordinate positions; and The physical length and physical width of each individual grid cell are set for the gridded resampling. The volume scalar value calculated by the above formula is generated and stored as the negative elevation integral. Simultaneously, the set of grid coordinates corresponding to the negative elevation region and the local relative height difference values ​​at each coordinate position are stored as negative elevation region mask data. This negative elevation region mask data retains the two-dimensional spatial index and corresponding relative elevation value of each grid cell for subsequent connected component searches.

[0049] Furthermore, the process of generating local tent deformation feature matrix data includes: extracting geometric feature vectors from particle contour data and normalizing the geometric feature vectors to generate normalized particle morphology data; mapping the film protrusion height data and the negative elevation integral according to a preset ratio to generate normalized spatial deformation data; and splicing the normalized particle morphology data and the normalized spatial deformation data in a tensor dimension according to a preset channel order, and aligning the spatial center based on the composite positioning point data to generate local tent deformation feature matrix data.

[0050] Read the solid-gas composite closed contour chain code from the particle contour data, calculate the two-dimensional projected area enclosed by the closed contour, the perimeter of the contour boundary, and the aspect ratio of the minimum bounding rectangle. Arrange these three geometric parameters in a fixed index order to construct a one-dimensional geometric feature vector. Obtain the theoretical maximum and minimum values ​​of each pre-set geometric parameter, and use the min-max normalization method to normalize the vector, unifying the values ​​of each dimension of the vector to a dimensionless range of 0 to 1. Then, establish a preset size (e.g., ...). A blank two-dimensional spatial grid (pixels) is used to map the closed contour chain code into the grid to generate a connected domain mask, and the normalized feature vector values ​​are assigned to the grid nodes of the mask coverage area by channel to generate a normalized particle morphology data matrix with the corresponding number of channels.

[0051] The system-preset nominal thickness constant and allowable micro-deformation volume critical constant of the heat-sealing film are obtained. The film protrusion height data and negative elevation integral are divided by the corresponding constants to perform dimensionless mapping, and the pure dimension height scaling factor and volume scaling factor are extracted. In a two-dimensional spatial grid of the same size as mentioned above, a two-dimensional Gaussian distribution attenuation model with the physical center of the grid as the origin is established. Specifically, the dimensionless height scaling factor is set as the central peak amplitude parameter of the Gaussian distribution model. For variance parameter Its physical meaning is the diffusion range of deformation influence in space (characterized by the standard deviation of a Gaussian distribution), which is pre-determined through the following calibration experiment: Select a standard spherical tungsten carbide particle with a known particle size (e.g., 0.5 mm), obtain the film protrusion height data and negative elevation integral of the particle, extract the measured local relative height difference matrix centered on the centroid, perform two-dimensional Gaussian surface least squares fitting on the measured height difference matrix, and use the fitted Gaussian variance... As Calibrate and calculate the corresponding volume ratio. Calibration, according to Calibration / The calibrated ratio determines the spatial diffusion coefficient β and is stored in the system as a preset parameter. For the same batch of heat-sealing film material, this calibration only needs to be performed offline once. During subsequent testing, the volume ratio of the current sample is multiplied by this pre-stored spatial diffusion coefficient. ,Right now .

[0052] The established two-dimensional Gaussian decay mapping formula: in, Local orthogonal coordinates with the physical center of the grid as the origin (unit: pixels, which can be converted to mm via the physical grid spacing); The central peak amplitude parameter is equal to the height scaling factor of the current sample (dimensionless, ranging from 0 to 1). The variance parameter is determined by the above calibration process (unit: mm²). All node coordinates within the grid are traversed and substituted into the Gaussian distribution attenuation model established above for numerical discretization, generating a two-dimensional gradient matrix with a center-to-outward radiation attenuation characteristic, which is used as the normalized spatial deformation data.

[0053] The composite positioning point data output in the previous steps is retrieved. Using the global physical coordinates of this composite positioning point data as the reference anchor point, the spatial translation offset of the grid center point of the normalized particle morphology data matrix and the normalized spatial deformation data matrix relative to this anchor point is calculated. Based on the calculated translation offset, bilinear interpolation translation transformation is performed on the two matrices to ensure that the spatial grid centers of both are strictly registered and aligned to the location of the composite positioning point data. After alignment, the registered normalized particle morphology data matrix and the normalized spatial deformation data matrix are concatenated and stitched along the channel dimension according to the preset tensor channel arrangement order. The multi-channel composite tensor data output after stitching is generated and stored as the local tent deformation feature matrix data. The local tent deformation feature matrix data is an M×M×C multi-channel composite tensor, where M is the preset two-dimensional spatial grid size and C is the total number of channels, which is determined by superimposing the number of channels of the normalized particle morphology data and the number of channels of the normalized spatial deformation data.

[0054] Furthermore, the process of outputting the predicted vacuum attenuation rate data includes: establishing a neighborhood graph structure for the negative elevation region mask data with a grid as the basic unit along the width direction from the inner to the outer side of the heat seal; performing a connected component search to extract the continuous recessed path that runs through the heat seal; calculating the channel morphology feature vector along the extension direction of the continuous recessed path to generate the cross-seal leakage channel vector data; inputting the cross-seal leakage channel vector data into the attenuation neural network for mapping, calculating and outputting the predicted vacuum attenuation rate data.

[0055] A three-dimensional neighborhood graph structure is established along the width direction from the inner edge to the outer edge of the heat-sealed area, using a grid as the basic unit. A preset deformation threshold is set, and grids in the negative elevation area mask data whose absolute relative elevation value is greater than this threshold are extracted and marked as passable nodes, with the corresponding values ​​used as node weights. A breadth-first search algorithm is applied to search for connected paths extending from the inner side to the outer side of the heat-sealed area in the three-dimensional neighborhood graph. Isolated recessed areas that fail to simultaneously contact the inner and outer boundaries of the heat seal are removed. A three-dimensional mesh connected set spanning the entire width of the heat seal is extracted as a continuous recessed path. Along the extension direction of the continuous recessed path, sections perpendicular to the channel centerline are cut at fixed step sizes. The product of the negative elevation difference value contained in each section and the physical area of ​​a single mesh is calculated and accumulated to obtain the cross-sectional shape integral of each section, with the unit being volume units. All cross-sectional shape integral values ​​within the entire length of the continuous recessed path are traversed, and the minimum value is extracted by bubble sort comparison and recorded as the minimum cross-sectional shape integral of the continuous recessed path.

[0056] Extract the three-dimensional geometric centerline of the continuous concave path in the spatial coordinate system, calculate the total actual spatial curve length of the centerline along the channel extension direction, and simultaneously extract the three-dimensional Euclidean spatial straight-line distance between the inner starting point and the outer ending point of the continuous concave path in the heat seal. Divide the total actual spatial curve length by the three-dimensional Euclidean spatial straight-line distance algebraically to calculate the tortuosity parameter reflecting the degree of channel curvature. Divide the extracted minimum cross-sectional morphology integral by the tortuosity parameter algebraically to calculate the channel flow capacity parameter (the larger the minimum cross-sectional morphology integral, the stronger the channel flow capacity; the larger the tortuosity, the weaker the channel flow capacity; this ratio serves as a proxy indicator reflecting the overall flow level of the channel). Simultaneously, the total physical length of the centerline of the continuous concave path and the average height difference within the channel are extracted. These three components characterize the geometric features of the leakage channel from different dimensions: the channel flow capacity parameter reflects the throughput capacity at the narrowest bottleneck of the leakage channel; the total physical length of the centerline reflects the cumulative friction resistance factor along the leakage path; and the average height difference, as a proxy variable for the channel cross-sectional depth, reflects the influence of the channel cross-sectional shape on the flow resistance. Together, they cover the bottleneck factor, friction factor, and depth factor describing the morphology of the concave path, providing a complementary input feature space for the subsequent attenuation neural network. The channel flow capacity parameter, the total physical length of the centerline, and the average height difference are arranged into a one-dimensional numerical matrix according to a fixed index order to generate the vector data of the cross-sealed leakage channel.

[0057] Multiple sets of offline historical sample data are acquired. Each set of sample data includes the cross-seal leakage channel vector extracted through the aforementioned steps, and the actual vacuum decay rate scalar value of the corresponding powder packaging bag obtained by the pressure difference vacuum decay tester. This data is used to construct a neural network training set, where the cross-seal leakage channel vector is used as the network input, and the actual vacuum decay rate scalar value is used as the supervision label. The decay neural network adopts a structure with an input layer receiving the channel vector, at least one fully connected hidden layer, and an output layer. The hidden layer uses a nonlinear activation function, and the output layer uses a sigmoid activation function to map the output to the 0 to 1 interval. During training, the mean squared error is used as the loss function, and the gradient descent algorithm is used for backpropagation to update the network weights. The loss is monitored on a preset validation set until the convergence condition is met, thus obtaining the pre-trained decay neural network. It should be noted that the vacuum decay rate depends not only on the statistical correlation between the surface morphology concavity path features and the internal sealing state, but also theoretically on factors such as the initial vacuum degree of the packaging bag, the internal volume of the bag, and the ambient temperature. However, under standardized continuous production conditions on the production line, the vacuum process parameters (target vacuum degree) of the same batch of packaging bags are kept consistent, the bag internal volume is fixed, and the ambient temperature fluctuates within the constant temperature control range of the workshop (e.g., 20±5℃). The impact of the batch variation of the above factors on the decay rate is much smaller than the impact caused by the difference in surface morphology proxy features. Therefore, the mapping relationship between the surface morphology proxy features and the vacuum degree decay rate learned by the decay neural network under the controlled conditions of the production line has sufficient engineering practical value. If the application scenario changes to packaging bags of different specifications or significantly different environmental conditions, the network can be retrained by changing the training dataset to adapt to the new conditions without changing the technical framework of this method. The generated cross-sealing leakage channel vector data is input into the pre-trained decay neural network, and the interval constraint mapping is performed through the activation function in the output layer to calculate and output the scalar form of the predicted vacuum degree decay rate data.

[0058] Furthermore, the process of outputting dynamic tear risk score data includes: extracting the local relative elevation matrix corresponding to the film protrusion height data; calculating the partial derivative of the local relative elevation matrix in the lifting force direction to generate force-direction elevation gradient matrix data; multiplying the equivalent tensile stress constant data of the target lifting load with the force-direction elevation gradient matrix data at corresponding positions to generate load-morphology gradient proxy matrix data; and inputting the load-morphology gradient proxy matrix data into a prediction neural network to calculate and output the dynamic tear risk score data.

[0059] The force direction parameters for lifting the packaging bag, set in the packaging production line, are obtained. This force direction vector in three-dimensional space is projected onto a two-dimensional grid plane containing the heat-sealing reference surface. The two-dimensional plane force direction vector of the lifting action on the film surface is extracted and decomposed into a horizontal component along the grid's horizontal axis and a vertical component along the grid's vertical axis. All two-dimensional grid coordinate nodes in the local relative elevation matrix are traversed. For each node, the elevation values ​​of adjacent nodes are extracted along the grid's horizontal and vertical axes. The first partial derivative of elevation along the horizontal axis and the second partial derivative of elevation along the vertical axis are calculated for each node using central difference. Subsequently, the first partial derivative of elevation is multiplied by the horizontal component, and the second partial derivative of elevation is multiplied by the vertical component. The two products are then added to calculate the rate of change of elevation along the force direction vector for each node. This represents the geometric steepness formed in the actual force direction after the internal rigid foreign object supports the film. The calculated rates of change of elevation for all nodes are arrayed and reorganized according to the row and column indices of the original two-dimensional grid coordinates to generate the force direction elevation gradient matrix data.

[0060] Simultaneously analyze the current working parameters, extract the dynamic lifting load data of the target packaging bag at the moment it is lifted and accelerated away from the workstation by the robotic arm, extract the measured lifting mass of the current packaging bag, the lifting acceleration of the robotic arm, and the gravitational acceleration, and synthesize the three into a dynamic lifting resultant force according to dynamic logic; simultaneously obtain the lateral width and nominal thickness of the heat-sealing film of the packaging bag and multiply them to obtain the material cross-sectional area of ​​the heat-sealing area, divide the dynamic lifting resultant force by the material cross-sectional area, and convert the macroscopic dynamic lifting load into equivalent tensile stress constant data acting on the unit area of ​​the heat-sealing area. Use this equivalent tensile stress constant data as a scalar multiplier and perform element-wise matrix scalar multiplication with the previously generated force-elevation gradient matrix data. Through this mapping operation, the gradient matrix representing the degree of geometric change on the film surface is coupled with the dynamic load base representing the macroscopic external force, so that the grid node with the larger directional elevation gradient receives a higher risk amplification weight, thereby generating load-morphology gradient proxy matrix data for relative risk ranking among samples. It should be noted that the above element-by-element proxy matrix calculation based on the equivalent tensile stress constant × elevation gradient employs an engineering simplification using the concept of stress concentration factor in linear elasticity. Heat-sealable films (such as polyethylene PE, polypropylene PP, or their multilayer composite films) are viscoelastic polymers, and their actual stress-strain relationship is time-dependent, exhibiting creep and stress relaxation behavior. However, under the standardized lifting operation conditions of the production line addressed by this method, this linear elastic simplification has the following rationale: First, the lifting action is completed instantaneously. The loading time for the robotic arm to grip the packaging bag and accelerate its removal from the workstation is typically no more than one second. At this short timescale, the polymer film mainly exhibits an instantaneous elastic response. The influence of viscous flow and stress relaxation on stress distribution is limited, and the instantaneous elastic modulus can effectively characterize the material's stiffness behavior. Second, in terms of deformation magnitude, the local strain caused by the tungsten carbide particles supporting the film usually falls within a small deformation range. The height of the film protrusion is on the order of hundreds of micrometers, while the lateral distribution scale of the heat-sealed film is on the order of millimeters—a difference of an order of magnitude. The local strain generally does not exceed 5%. At this strain level, neither geometric nonlinearity nor material nonlinearity is significant, and the assumption of linear elastic small deformation is valid. Third, and more importantly, the goal of this method is not to perform precise strength verification or calculate the exact megapascals of stress at a specific location. Instead, it aims to generate a scoring matrix that reflects the relative strength ranking of stress concentrations in different local areas of the heat seal. This matrix serves as the feature input for a predictive neural network to assess tear risk. The linear elastic simplification is sufficiently accurate in capturing the location distribution and relative strength ranking of stress concentrations, making it highly valuable in engineering applications. In summary, under the constraints of standardized lifting operations on a production line, simplifying the mechanical behavior of the film into a linear elastic model to generate the load-morphology gradient surrogate matrix is ​​a reasonable engineering approximation, striking a proper balance between computational efficiency and physical accuracy.

[0061] The next step is the training phase of the predictive neural network. First, multiple sets of historical sample data are collected offline to construct a training set. The input samples are load-morphology gradient proxy matrix data generated using hard particles of different sizes according to the aforementioned process. How are labels created? Each sample packaging bag is placed on a tensile testing machine for destructive tensile testing, and the peak ultimate tensile force at which the film undergoes substantial tearing is measured. Then, the target dynamic lifting force is divided by this peak ultimate tensile force to obtain the load utilization rate label for that sample. During training, the tear risk prediction neural network adopts a structure where the input layer receives the load-morphology gradient proxy matrix, alternating layers of stacked two-dimensional convolutional and max-pooling layers, a fully connected backend layer, and an output layer. The convolutional layers use nonlinear activation functions, and the output layer outputs a scalar risk score after mapping through the activation function. The mean squared error is used as the loss function, and the gradient descent algorithm is used to update the network weights through backpropagation. The loss is monitored on a preset validation set until the convergence condition is met, resulting in a pre-trained tear risk prediction neural network.

[0062] In actual testing, the load-morphology gradient proxy matrix data is fed into the trained prediction neural network as a single-channel two-dimensional tensor. The front-end convolutional kernel slides across the matrix, automatically capturing local gradient abrupt changes. Next, a max-pooling layer performs downsampling, filtering out secondary information and focusing the network on high-risk response regions. At the back end, a fully connected layer flattens the extracted high-dimensional gradient abrupt changes, redistributes the weights, and finally, after passing through a non-linear activation function in the output layer, outputs a continuous scalar value. This is the dynamic tear risk score of the heat seal under the current stress condition.

[0063] This rating ranges from 0 to 1. A value closer to 0 indicates a lower risk of tearing at the heat seal under the current stress conditions; a value closer to 1 indicates a higher risk. It's important to clarify that this rating measures the relative ranking of tear risk among different heat-sealed samples. In other words, a sample with a rating of 0.8 is indeed more likely to fail under lifting conditions than a sample with a rating of 0.4, but it is not a statistically precise probability of breakage. In practical engineering applications, simply comparing it to a preset safe tear rating threshold (e.g., 0.15) is sufficient to determine whether the seal is acceptable or unacceptable.

[0064] The final step in quality assessment is to compare two predicted results—the predicted vacuum attenuation rate and the dynamic tear risk score—with preset thresholds in the system. These thresholds are pre-calibrated by technicians based on the material characteristics of the target packaging bag and the factory quality inspection standards, and are directly accessed by the system. The comparison logic is an "OR" relationship: if the predicted vacuum attenuation rate exceeds the allowable vacuum attenuation rate threshold, or the dynamic tear risk score exceeds the safe tear risk score threshold, if either indicator exceeds the standard, the current heat seal is deemed to have a risk of sealing failure or stress damage, and a non-conforming label is generated. Only if both predicted data do not exceed their respective thresholds is the heat seal's physical strength and sealing performance deemed to meet the standards, and a conforming label is generated. Finally, the quality assessment label is sent to the production line control system.

[0065] It should be noted that the following abnormal operating conditions may occur during actual production line operation, and this method should be handled according to the following branches when implemented. When no tungsten carbide particles are detected in the visible light transmission image, it is determined that there are no hard inclusions in the current heat-sealed area, and the morphology extraction and network inference steps are skipped directly, and a qualified label is output. When multiple tungsten carbide particles are detected in a single heat-sealed area, the above steps are performed independently for each particle to generate its own composite positioning point data, local tent deformation feature matrix, cross-sealing leakage channel vector, and load-morphology gradient proxy matrix. The maximum predicted vacuum attenuation rate and the maximum dynamic tear risk score among all particles are taken as the final evaluation indicators for the heat-sealed area. When the cumulative occlusion area of ​​tungsten carbide particles in the transmission image exceeds the preset upper limit, or when the structured light phase unpacking fails and a large area of ​​point cloud is missing, the current heat-sealed area is marked as invalid, triggering the production line re-inspection mechanism, and no qualified or unqualified label is output. When the connected region search does not find any continuous concave path from the inside to the outside of the heat-sealed area, each component in the cross-sealing leakage channel vector is set to zero, and the attenuation neural network outputs the corresponding predicted vacuum attenuation rate accordingly, and the threshold comparison process is entered normally. When the input feature value of the model exceeds the distribution range of the training set, a low confidence label is attached to the output risk score, triggering the production line to manually re-examine the sample.

[0066] Example 2: The current production line uses heavy-duty drums and bags to package tungsten carbide powder. The process involves two independent heat sealing steps: first, sealing the bottom, then filling 50KG of tungsten carbide powder, and finally vacuuming and sealing the top. The visual inspection system of this invention is deployed at both the upper sealing station and the lower sealing station.

[0067] On the upper sealing side, the focus is on airtightness. After heat sealing and vacuuming are completed, a heat sealing completion signal is received, and visible light transmission images and surface depth point cloud images of the upper heat-sealed area are acquired simultaneously. In the step of generating composite positioning point data, the weighting coefficients are biased towards the geometric center of the micro-cavity area to enhance the capture of potential weak sealing areas. Then, negative elevation area mask data is extracted to search for continuous recessed paths that penetrate the heat seal. The geometric features of the channels are used to create a cross-seal leakage channel vector, which is then fed into an attenuation neural network to calculate the predicted vacuum attenuation rate of the upper seal.

[0068] On the bottom sealing side, after 50KG of powder has been filled, the bottom seal bears an extremely high load, making its stability paramount. After the bottom heat seal is completed, multimodal images of this area are also acquired. If tungsten carbide hard particles are trapped in the bottom seal area during filling, the film protrusion height data is first accurately extracted using anisotropic differential operators and surface fitting algorithms, and then its elevation gradient in the lifting force direction is calculated to generate a force-direction elevation gradient matrix. Simultaneously, the system analyzes the extreme dynamic load parameters when lifting a 50KG bag and converts them into the equivalent tensile stress constant of the target lifting load. This constant is then multiplied element-wise by the force-direction elevation gradient matrix at corresponding positions to obtain the load-morphology gradient proxy matrix. After being fed into the predictive neural network, the network calculates and outputs a dynamic tear risk score to assess whether the bottom seal area will rupture and leak powder under the extreme lifting condition of 50KG.

[0069] Finally, the system compares the predicted vacuum attenuation rate of the upper seal and the dynamic tear risk score of the lower seal with their corresponding preset thresholds. If both exceed the limits, the system determines that the 50KG heavy-duty drum bag meets the physical strength and sealing performance standards and issues a qualified label. If either exceeds the limit, it issues an unqualified label and is directly rejected by the production line. It should be noted that the above bag-level judgment rule treats the sealing risk of the upper seal area and the lifting damage risk of the lower seal area as two independent dimensions, and aggregates them using a logical OR relationship: the packaging bag is judged to be qualified only when the risk indicators of both heat-sealed areas do not exceed the corresponding thresholds; if either area exceeds the limit, it is judged to be unqualified. For packaging bag types with only one heat seal, the judgment result of that single heat-sealed area is directly used as the bag-level label.

[0070] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based visual inspection method for powder packaging quality, characterized in that, include: Receive the heat sealing completion signal and acquire visible light transmission image and surface depth point cloud image of the heat-sealed area; The visible light transmission image is input into the anisotropic differential operator to extract particle contour data and composite positioning point data; the composite positioning point data is mapped to the surface depth point cloud image, and local depth matrix data is extracted according to preset cropping parameters. The local depth matrix data is input into the surface fitting algorithm to calculate the maximum normal height difference. The local depth matrix is ​​then subtracted element by element from the fitting reference surface to obtain the local relative elevation matrix. From the local relative elevation matrix, the film protrusion height data, the mask data of the negative elevation region, and the negative elevation integral are extracted. The particle contour data, the film protrusion height data, and the negative elevation integral are then concatenated to generate the local tent deformation feature matrix data. A connected region search is performed on the mask data of the negative elevation region along the sealing width direction to extract the continuous path through the heat seal and calculate the leakage channel vector data across the seal. The data is then input into the attenuation neural network and outputs the predicted vacuum attenuation rate data. The equivalent tensile stress constant data of the target lifting load and the elevation gradient along the force direction of the film protrusion height data are fused to generate load-morphology gradient proxy matrix data. This data is then input into the prediction neural network and outputs the dynamic tear risk score data. The predicted vacuum decay rate data and dynamic tear risk score data are compared with preset thresholds to output the quality judgment label data of the heat-sealed area.

2. The method for visual inspection of powder packaging quality based on deep learning according to claim 1, characterized in that, The process of receiving the heat-sealing completion signal and acquiring the visible light transmission image and surface depth point cloud image of the heat-sealed area includes: receiving the heat-sealing completion trigger signal sent by the main control device; responding to the heat-sealing completion trigger signal, controlling the transmission light source located below the heat-sealed area to emit light, and controlling the first camera located directly above the heat-sealed area to acquire the light penetrating the heat-sealed area, generating the visible light transmission image data; synchronously controlling the structured light generator located above the heat-sealed area to project a structured light pattern onto the heat-sealed area, controlling the second camera arranged at a preset relative angle to the structured light generator to acquire the image of the heat-sealed area projected with the structured light pattern, and performing three-dimensional reconstruction based on the acquired image to generate the surface depth point cloud image data.

3. The method for visual inspection of powder packaging quality based on deep learning according to claim 1, characterized in that, The process of generating particle contour data and composite positioning point data includes: inputting the visible light transmission image into an anisotropic differential operator based on the direction of the airflow line, extracting the absolute extinction region formed by the tungsten carbide hard core, and the annular micro-cavity region formed by the micro-airflow blockage around the absolute extinction region, generating solid-gas composite inclusion contour data as the particle contour data; calculating the convex hull polygon of the absolute extinction region, extracting the coordinates of the sharp vertex of the polygon with the largest convex distance along the heat-sealing stretching direction; and vector superimposing the coordinates of the sharp vertex of the polygon with the geometric center coordinates of the annular micro-cavity region to generate the composite positioning point data.

4. The deep learning-based visual inspection method for powder packaging quality according to claim 1, characterized in that, The process of extracting local depth matrix data according to preset truncation parameters includes: mapping the composite positioning point data to the surface depth point cloud image according to camera calibration parameters to locate the three-dimensional coordinates of the deformation vertex; extracting the heat-sealing embossing direction vector within a preset range around the three-dimensional coordinates of the deformation vertex; extracting a first elevation attenuation gradient along the direction of the heat-sealing embossing direction vector, and extracting a second elevation attenuation gradient along the direction perpendicular to the direction of the heat-sealing embossing direction vector; using the positions where the first elevation attenuation gradient suddenly becomes zero and the positions where the second elevation attenuation gradient suddenly becomes zero as boundary points to construct eccentric elliptical boundary data; and truncating local point clouds from the surface depth point cloud image along the eccentric elliptical boundary data to generate the local depth matrix data.

5. The method for visual inspection of powder packaging quality based on deep learning according to claim 1, characterized in that, The process of generating film protrusion height data, negative elevation region mask data, and negative elevation integral includes: inputting local depth matrix data into a least-squares thin-plate spline fitting algorithm to construct heat-sealing reference surface data that fits the undeformed base surface of the heat-sealing film; performing matrix subtraction between the local depth matrix data and the heat-sealing reference surface data to generate local relative height difference matrix data; extracting the maximum positive value from the local relative height difference matrix data to generate film protrusion height data; extracting the set of grid coordinates and their relative elevation values ​​of negative elevation regions located around the film protrusion height data and with values ​​less than zero from the local relative height difference matrix data to generate negative elevation region mask data; performing integral calculation along the normal of the heat-sealing reference surface on the negative elevation region, and using the cumulative value of the product of the absolute value of the relative elevation within the negative elevation region and the physical area of ​​the grid as the negative elevation integral.

6. The deep learning-based visual inspection method for powder packaging quality according to claim 1, characterized in that, The process of generating local tent deformation feature matrix data includes: extracting geometric feature vectors from particle contour data and normalizing the geometric feature vectors to generate normalized particle morphology data; mapping the film protrusion height data and the negative elevation integral according to a preset ratio to generate normalized spatial deformation data; and splicing the normalized particle morphology data and the normalized spatial deformation data in a tensor dimension according to a preset channel order, and aligning the spatial center based on the composite positioning point data to generate local tent deformation feature matrix data.

7. The method for visual inspection of powder packaging quality based on deep learning according to claim 1, characterized in that, The process of outputting the predicted vacuum attenuation rate data includes: establishing a neighborhood graph structure for the negative elevation region mask data with a grid as the basic unit along the width direction from the inner to the outer side of the heat seal; performing a connected component search to extract the continuous recessed path that runs through the heat seal; calculating the channel morphology feature vector along the extension direction of the continuous recessed path to generate the cross-seal leakage channel vector data; inputting the cross-seal leakage channel vector data into the attenuation neural network for mapping, calculating and outputting the predicted vacuum attenuation rate data.

8. The deep learning-based visual inspection method for powder packaging quality according to claim 1, characterized in that, The process of outputting dynamic tear risk score data includes: extracting the local relative elevation matrix corresponding to the film protrusion height data; calculating the partial derivative of the local relative elevation matrix in the lifting force direction to generate force-direction elevation gradient matrix data; multiplying the equivalent tensile stress constant data of the target lifting load with the force-direction elevation gradient matrix data at corresponding positions to generate load-morphology gradient proxy matrix data; and inputting the load-morphology gradient proxy matrix data into a prediction neural network to calculate and output the dynamic tear risk score data.

Images