Intelligent box sealing control method based on multi-dimensional self-adaptive adjustment

Abstract

The invention relates to the technical field of intelligent box sealing, in particular to an intelligent box sealing control method based on multi-dimensional adaptive adjustment, and the method comprises the steps: carrying out the multi-modal feature fusion of RGB images collected by an industrial camera and depth data (three-dimensional point cloud, normal vectors and pose parameters), and combining the cavity convolution texture analysis and three-dimensional topology modeling, thereby achieving the intelligent box sealing control. According to the method, self-adaptive adjustment parameters used for adjusting the width of a carton conveying channel and the height of an adhesive tape pressing mechanism are generated by fusing surface textures (scratches and creases) and a three-dimensional structure (curvature and normal vectors); the equipment can adapt to complex working conditions such as special-shaped box bodies, printing pattern interference and surface deformation in real time, can sense abnormity in real time and automatically adjust, prevents inclined adhesive tape pasting, and thus adapts to complex and changeable actual production environments.

Description

Technical Field

[0001] The present invention relates to the field of intelligent carton sealing technology, and in particular to an intelligent carton sealing control method based on multi-dimensional adaptive adjustment. Background Art

[0002] At present, the process of automatically sealing cartons through sealing machines is a key link in modern packaging production lines, which mainly relies on mechanization and automation technology.

[0003] During the automatic sealing process of cartons, existing technologies mostly rely on a single type of sensor (such as photoelectric switches or 2D vision) to perform visual inspection of cartons. However, photoelectric switches can only determine whether the carton is in place and cannot obtain size, shape or surface features. Moreover, it is impossible to accurately analyze carton deformation, surface texture and other features through 2D images alone. 2D images cannot detect deformation defects such as warping and dents of cartons. 2D images only locate edges through grayscale or color contrast and are easily interfered with. Color printed patterns, reflective logos or stripes on the surface of the carton are also likely to interfere with edge detection of the carton (edge ​​detection is used to locate the edge of the carton so that the tape can be correctly applied). Misjudgments are easily made during the taping process, resulting in skewed or loose tape.

[0004] Therefore, there is a need for an intelligent carton sealing control method based on multi-dimensional adaptive adjustment that can solve the above problems. Summary of the Invention

[0005] The present invention provides an intelligent carton sealing control method based on multi-dimensional adaptive adjustment. The present invention realizes high-precision dynamic perception of carton geometric dimensions, surface defects and posture inclination by integrating multimodal feature fusion of RGB images + depth data (3D point cloud, normal vector, posture parameters) collected by industrial cameras, combined with void convolution texture analysis and 3D topological modeling. The method generates adaptive adjustment parameters for adjusting the width of the carton conveying channel and the height of the tape pressing mechanism by fusing surface texture (scratches, creases) and 3D structure (curvature, normal vector), so that the equipment can adapt to complex working conditions such as special-shaped boxes, printed pattern interference, surface deformation, etc. in real time. It can instantly perceive abnormalities and automatically adjust to prevent crooked tape, thereby adapting to the complex and changeable actual production environment.

[0006] The technical solution adopted by the present invention to solve the above technical problems is: an intelligent box sealing control method based on multi-dimensional adaptive adjustment, comprising the following steps:

[0007] Step 1: First, guide the carton to the carton sealing machine's infeed station. When the carton triggers the infeed sensor, an industrial camera simultaneously captures the carton's RGB image and depth data. The depth data includes a depth map, 3D point cloud data, normal vector data, confidence map, and pose parameters. The carton's geometric boundaries and pose parameters are extracted using a 3D reconstruction algorithm.

[0008] Step 2: The visual data is fed into a multimodal feature fusion algorithm, which combines the carton surface texture features with the 3D topology to generate adaptive adjustment parameters for adjusting the width of the carton conveyor channel and the height of the tape pressing mechanism.

[0009] Step 3: The conveyor belt guide plate spacing is driven according to the adaptive adjustment parameters, the width of the carton conveying channel is adjusted, and the height of the tape pressing mechanism is adjusted to seal the carton. At the same time, the tape rewind motor is dynamically controlled based on the tape deformation feedback to adjust the tape tension.

[0010] Step 4: When the labeling sensor is triggered, the conveyor belt stops and a cutting path is generated through sub-pixel edge detection. The cutter mechanism is driven to cut the tape. At the same time, the tape is pressed again by the tape pressing mechanism to strengthen the tape bonding strength and complete the box sealing.

[0011] Step 5: At the moment of cutting the tape, the temperature of the tape end is monitored in real time by an infrared thermal imager. If the temperature is lower than the viscosity activation threshold, the auxiliary heater is activated to pulse heat the cutting point.

[0012] Step 6: After the carton is sealed, the conveyor belt is restarted to output the carton. Sealing defects are identified based on multispectral imaging and defect detection algorithms, and process optimization parameters are generated and iterated to the next work cycle.

[0013] Furthermore, the extraction of the carton geometric boundary and posture parameters based on the three-dimensional reconstruction algorithm in step 1 includes:

[0014] Step 1-1: Perform bilateral filtering preprocessing on the depth map collected by the industrial camera to retain edge information while suppressing noise;

[0015] Step 1-2: Use the random sampling consistency algorithm to fit the plane equations of the top and side surfaces of the carton and remove outliers;

[0016] Step 1-3: Calculate the plane intersection line as the preliminary geometric boundary;

[0017] Steps 1-4: Fusing normal vector data to optimize boundary curvature continuity;

[0018] Steps 1-5: Perform weighted correction on the boundary points according to the confidence map to generate the final geometric model.

[0019] Furthermore, in step 2, the visual data is input into a multimodal feature fusion algorithm, and the adaptive adjustment parameters for adjusting the width of the carton conveying channel and the height of the tape pressing mechanism are generated by combining the carton surface texture features and the three-dimensional topological structure. The parameters include:

[0020] Step 2-1: Perform gamma correction and histogram equalization preprocessing on the carton RGB image;

[0021] Step 2-2: Extract local texture features of scratches, creases, and printed patterns on the carton surface through dilated convolution;

[0022] Step 2-3: Construct a topological structure based on the 3D point cloud, aggregating the curvature and normal vector differences of adjacent points;

[0023] Step 2-4: Perform channel attention weighting on texture features and topological features;

[0024] Step 2-5: Fusion of texture features and topological features through a gating mechanism to output adaptive adjustment parameters;

[0025] Step 2-6: Perform L2 regularization on the fusion parameters to prevent overfitting.

[0026] Furthermore, the control method of dynamically controlling the tape rewinding motor based on tape deformation feedback in step 3 includes:

[0027] Step 3-1: Deploy a high-speed polarization camera above the tape bonding area to capture the light intensity distribution on the tape surface at 1200 fps;

[0028] Step 3-2: Calculate the local strain energy density based on the light intensity gradient change;

[0029] Step 3-3: When the strain energy density exceeds the material yield threshold, an audible and visual alarm is triggered;

[0030] Step 3-4: Calculate the tension compensation value through the fuzzy controller;

[0031] Step 3-5: Dynamically adjust the speed of the tape rewinding motor using the PID algorithm;

[0032] Step 3-6: Update the control command every 50ms to form a closed-loop feedback.

[0033] Furthermore, generating a cutting path by sub-pixel edge detection in step 4 includes:

[0034] Step 4-1: Construct a Gaussian pyramid in the trigger area of ​​the labeling sensor and perform 3-level downsampling;

[0035] Step 4-2: Use the Sobel operator to extract the rough positioning area of ​​the tape edge;

[0036] Step 4-3: Calculate the sub-pixel edge offset based on the Zernike moment;

[0037] Step 4-4: Generate a smooth cutting path by fitting a cubic Bezier curve;

[0038] Step 4-5: Use the inverse kinematics model to convert the path into servo motor pulse instructions;

[0039] Step 4-6: Start the pressing roller 0.1 second before cutting and apply a dynamic pressure of 0.8-1.2 MPa.

[0040] Furthermore, starting the auxiliary heater to pulse-heat the cutting point in step 5 includes the following sub-steps:

[0041] Step 5-1: Use an infrared thermal imager to obtain the temperature matrix of a 5mm×5mm area at the cutting point;

[0042] Step 5-2: Calculate the variance of the temperature gradient distribution. If the variance exceeds the threshold, it is determined to be a local low temperature.

[0043] Step 5-3: Start the auxiliary heater to compensate for the low temperature area;

[0044] Step 5-4: Use PID-PWM controller to adjust the heating power;

[0045] Step 5-5: Collect temperature feedback every 10ms until the temperature stabilizes within the set range;

[0046] Step 5-6: If compensation fails three times in a row, a shutdown and maintenance instruction is triggered.

[0047] Furthermore, in step 6, identifying the sealing defects based on the multispectral imaging and defect detection algorithm and generating the process optimization parameters to iterate to the next working cycle include:

[0048] Step 6-1: Detect the gap between the edges of the tape using visible light and mark the continuous area exceeding 0.3 mm.

[0049] Step 6-2: Analyze the uniformity of adhesive layer penetration in the near-infrared band and calculate the effective contact area ratio;

[0050] Step 6-3: Fuse multispectral feature maps through feature pyramid network;

[0051] Step 6-4: Perform morphological opening operation filtering on the bubble defect to eliminate artifact interference;

[0052] Step 6-5: Generate a heat map based on defect type and location;

[0053] Step 6-6: Optimize the tension, speed, and temperature parameter combinations using the gradient descent algorithm;

[0054] Step 6-7: Encrypt the optimized parameters and write them into the process parameter queue of the non-volatile memory;

[0055] Steps 6-8: When the next box sealing cycle starts, extract the latest parameters from the queue and verify their validity;

[0056] Step 6-9: If the parameter verification passes, the set values ​​of tape tension, conveying speed and heating temperature are dynamically updated;

[0057] Step 6-10: If the parameter verification fails, the historical parameter set is enabled and the parameter self-learning module is triggered.

[0058] Furthermore, the gating mechanism in steps 2-5 includes:

[0059] Step 2-5-1: Perform spatial attention weighting on the texture feature map to highlight the scratch area;

[0060] Step 2-5-2: Perform channel attention weighting on the topological feature map to strengthen the curvature mutation points;

[0061] Step 2-5-3: Modeling feature temporal dependencies through gated recurrent units;

[0062] Step 2-5-4: Perform batch normalization on the fused features.

[0063] Furthermore, the calculation of the tension compensation value by the fuzzy controller in step 3-4 includes the following steps:

[0064] Step 3-4-1: Define strain error and error change rate as fuzzy input variables;

[0065] Step 3-4-2: Build a fuzzy rule base to cover all working condition combinations;

[0066] Step 3-4-3: Use the centroid method to defuzzify and output the compensation value;

[0067] Step 3-4-4: Perform dead zone compensation on the output value to prevent high-frequency oscillation.

[0068] Furthermore, generating a smooth cutting path by fitting a cubic Bezier curve in step 4-4 includes:

[0069] Step 4-4-1: Select path control points at equal intervals along the tape edge based on the sub-pixel positioning results, and record the coordinates and normal vector directions of each point;

[0070] Step 4-4-2: Construct a curvature continuity constraint equation to ensure that the curvature change rate between adjacent control points does not exceed 5%;

[0071] Step 4-4-3: Use the particle swarm optimization algorithm to iteratively calculate the spatial coordinates of the control points. The objective function is to minimize the total curvature change rate of the path.

[0072] Step 4-4-4: Perform quintic spline interpolation on the optimized control points to generate a cubic Bezier curve path with continuous curvature;

[0073] Step 4-4-5: Discretize the curve path into 1000 interpolation points and convert them into the pulse sequence of the servo motor through the inverse kinematics model;

[0074] Step 4-4-6: Verify the path smoothness in the simulation environment. If the maximum acceleration exceeds 1.5g, return to step 4-4-3 and re-optimize.

[0075] The advantages of the present invention are:

[0076] The present invention achieves high-precision dynamic perception of carton geometric dimensions, surface defects and posture inclination by integrating multimodal feature fusion of RGB images + depth data (3D point cloud, normal vector, posture parameters) collected by industrial cameras, combined with void convolution texture analysis and 3D topological modeling. This method generates adaptive adjustment parameters for adjusting the width of the carton conveying channel and the height of the tape pressing mechanism by fusing surface texture (scratches, creases) with 3D structure (curvature, normal vector), so that the equipment can adapt to complex working conditions such as special-shaped boxes, printed pattern interference, surface deformation, etc. in real time. It can instantly sense anomalies and automatically adjust to prevent crooked tape, thereby adapting to the complex and changeable actual production environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0078] Figure 1 This is a flow chart of an intelligent carton sealing control method based on multi-dimensional adaptive adjustment provided by the present invention.

[0079] Figure 2 Schematic diagram of the three-dimensional structure of the present invention Figure 1 ;

[0080] Figure 3 Schematic diagram of the three-dimensional structure of the present invention Figure 2 ;

[0081] in:

[0082] 1. Carton sealing machine; 2. Carton infeed stopper; 3. Carton infeed sensor switch;

[0083] 4. Conveyor belt; 5. Adhesive tape; 6. Adhesive tape rewinding motor;

[0084] 7. Adhesive tape motor; 8. Labeling sensor switch. DETAILED DESCRIPTION

[0085] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0086] Example 1:

[0087] Figure 1 This is a flow chart of an intelligent carton sealing control method based on multi-dimensional adaptive adjustment provided by the present invention. Figure 2 Schematic diagram of the three-dimensional structure of the present invention Figure 1 , Figure 3 Schematic diagram of the three-dimensional structure of the present invention Figure 2 ,like Figure 1 , Figure 2 and Figure 3 The intelligent carton sealing control method based on multi-dimensional adaptive adjustment shown in the figure includes the following steps:

[0088] Step 1: Carton 3D geometric parameter extraction and pose estimation, triggering carton infeed and data collection. Carton infeed trigger mechanism: When a carton enters the carton sealer's infeed station, the infeed sensor switch (model: Omron E2B-M12KN08-WP-B1) is activated. This switch utilizes a non-contact photoelectric principle, with a detection range of 8 mm and a response time of less than 1 millisecond, ensuring an immediate trigger signal upon carton arrival. The trigger signal is transmitted via a PLC (Siemens S7-1200) to the industrial camera and control system, initiating the synchronous data collection process.

[0089] Industrial camera configuration: A Basler ace2 acA2440-75uc industrial camera is installed directly above the carton loading station, 500 mm from the carton surface, with a field of view covering 600 mm × 400 mm.

[0090] The camera has a resolution of 2448×2048 pixels, a frame rate of 30 frames per second, is equipped with a global shutter, and supports external trigger synchronous acquisition. Data acquisition content includes RGB images: acquiring color images and depth data of the carton surface for analyzing printed patterns, scratches and surface textures. Depth data includes: Depth map: records the physical distance of each pixel with an accuracy of ±0.5 mm. Three-dimensional point cloud: generates three-dimensional coordinate data of the carton surface with a point density of 10 points per square millimeter. Normal vector: calculates the normal direction of each point for surface curvature and deformation analysis. Confidence map: marks the reliability of depth data (0-1 range), and low confidence areas (<0.8) require subsequent correction. Posture parameters: including the coordinates of the center of the carton (X, Y, Z), tilt angle (α, β) and rotation angle (γ).

[0091] The 3D reconstruction algorithm includes steps 1-1: Depth map preprocessing: bilateral filtering, using both spatial and grayscale domain filtering to balance noise suppression and edge preservation. A spatial kernel size of 5×5 pixels is used to smooth local noise. A grayscale standard deviation of 0.1 is used to suppress outliers caused by surface reflections or shadows. After filtering, the noise standard deviation of the depth map is reduced from ±1.2 mm to ±0.3 mm, with less than 5% loss in edge sharpness.

[0092] Step 1-2: Plane fitting and outlier removal: The RANSAC algorithm is configured as follows:

[0093] The maximum number of iterations is 1000, ensuring a high probability of finding the optimal plane.

[0094] The internal point determination threshold is 0.5 mm. Points that deviate from the plane by more than this value are considered outliers.

[0095] Minimum inlier ratio: 80%. If it is lower than this ratio, resampling will be performed.

[0096] The process involves randomly selecting three points from the depth map to generate an initial plane hypothesis. The distances of all points to the plane are calculated, and the number of inliers that meet a threshold is counted. This process is repeated until the plane with the most inliers is found.

[0097] For the standard cubic carton test, the fitting errors of the top and side planes are ±0.2 mm and ±0.3 mm respectively, and the outlier rejection rate is about 12%.

[0098] Steps 1-3: Geometric Boundary Calculation: Planar intersection calculations are performed by solving the top and side plane equations simultaneously. For example, if the top equation coefficients are 0.982, 0.141, and 0.124, and the side equation coefficients are 0.707, -0.707, and 0.001, the intersection lines form the carton edges. Boundary Generation: Extract the carton edges along the intersection lines as the preliminary geometric boundary, marking the intersection line between the top and side surfaces.

[0099] Steps 1-4: Curvature Continuity Optimization: Calculate the average normal vector for each boundary point and its 10 adjacent normals. Adjust the boundary point position to ensure the curvature change between adjacent points is ≤3%. This process involves scanning point by point along the boundary line and calculating the angle between the normal vectors of the current point and the previous point. If the angle is greater than 5°, adjust the current point coordinates to ensure curvature continuity.

[0100] Steps 1-5: Confidence weighted correction:

[0101] Weight calculation: Assign a weight to each boundary point, and the weight value is 70% of the confidence value.

[0102] The correction rules include: if the confidence level is less than 0.8, use the adjacent high confidence points for interpolation correction; if the confidence level is greater than or equal to 0.8, retain it directly.

[0103] Output results: Final geometric model accuracy is ±0.2 mm, tilt angle error is <0.5°, and rotation angle error is <1°.

[0104] Step 2: Multimodal feature fusion generates adjustment parameters:

[0105] Step 2-1: RGB image preprocessing:

[0106] Gamma Correction: A correction factor of 2.2 was set to enhance shadow detail and avoid overexposure in highlights. Histogram Equalization: CLAHE (Contrast-Limited Adaptive Histogram Equalization) was used, with a block size of 32×32 pixels and a contrast limit threshold of 2.0 to prevent local overenhancement. The processed image's signal-to-noise ratio increased by 8dB, and the contrast between scratches and printed patterns was enhanced by 40%.

[0107] Step 2-2: Surface texture feature extraction:

[0108] Dilated Convolutional Network: A network with a dilation rate of 2 and a convolution kernel of 3×3 is used to extract scratches, creases, and printed patterns on the surface of the cardboard.

[0109] The feature extraction process includes the following: A 224×224 pixel RGB image is input. After three layers of dilated convolution and pooling, a 112×112 pixel feature map is output. Scratches with a width of ≥ 0.3 mm can be detected with a positioning accuracy of ±0.1 mm.

[0110] Step 2-3: 3D topology modeling:

[0111] Topology Construction: Based on 3D point cloud data, adjacent points are connected with a 5mm neighborhood radius to construct a topology structure. Feature Calculation: Areas with a curvature greater than 0.15 are marked as concave. Areas with a difference angle greater than 10° are marked as warped.

[0112] Steps 2-4: Channel-wise attention weighting applies spatial weighting to the texture feature map, highlighting scratches and creases. Channel-wise weighting is applied to the topological feature map, enhancing the features of curvature mutation points. Attention weights are generated through global average pooling and fully connected layers.

[0113] Steps 2-5: Gating Mechanism Fusion: For the GRU unit, texture features and topological features are input, and timing information is dynamically integrated through update gates and reset gates. Output parameters include the width adjustment of the transmission channel (dynamically adjusting the guide plate spacing) and the height compensation value of the pressing mechanism (compensating for deformation error).

[0114] Step 2-6: Regularization processing: impose regularization constraints on the fusion parameters to prevent overfitting.

[0115] Step 3: Dynamic closed-loop control of tape tension:

[0116] Step 3-1: High-speed polarization intensity acquisition: The camera used is a FLIR BFS-U3-200S6M-C high-speed polarization camera with a frame rate of 1200 fps. It is installed above the tape bonding area (external device, not shown in the figure) at a 45° angle to the tape plane.

[0117] Polarizer configuration: The direction of the linear polarizer is consistent with the stretching direction of the tape to enhance strain sensitivity.

[0118] Step 3-2: Strain energy density calculation: Calculate the local tensile strain based on the change in polarized light intensity. When the strain energy density exceeds the yield threshold of the polypropylene tape (0.1 joules / cubic meter), the sound and light alarm (Honeywell 9450) is triggered and the abnormal location is recorded.

[0119] Step 3-3: Fuzzy rule base construction: Input variables: strain error (classified as large, medium, and small) and error change rate (positive, zero, and negative). Output variable: tension compensation value (classified as seven levels: positive large, medium, and negative). If the strain error is large and the error change rate is positive, the output is a positive large compensation value.

[0120] Step 3-4: Dynamically adjust the motor speed: PID parameters: Proportional coefficient 0.8, integral time 0.1 second, differential time 0.05 second. Control cycle: Update the tape rewind motor speed every 50 milliseconds, and control tension fluctuation within ±3 Newtons.

[0121] Step 4: Sub-pixel cutting path planning and execution:

[0122] Step 4-1: Perform three-level downsampling on the original image (2448×2048 pixels) of the labeling trigger area to generate a low-resolution image of 307×256 pixels to accelerate the coarse edge positioning.

[0123] Step 4-2: Use the Sobel operator to calculate the horizontal and vertical gradients, set the gradient threshold to 50, and filter out low-contrast areas.

[0124] Step 4-3: Calculate the sub-pixel edge offset based on the Zernike moment to improve the edge positioning accuracy to ±0.02 mm.

[0125] Step 4-4: Select 9 control points at equal intervals along the sub-pixel edge, with a spacing of 5 mm. Limit the curvature change rate between adjacent control points to within 5% to ensure a smooth path.

[0126] Steps 4-5: Discretize the Bezier path into 1000 interpolation points and convert them into pulse instructions for the servo motor (Yaskawa SGM7G-20A), with a pulse equivalent of 0.001 mm.

[0127] Steps 4-6: Automatically select the pressure value (0.8-1.2 MPa) according to the tape material (such as BOPP or hot melt adhesive), and start the pressing roller 0.1 second before cutting.

[0128] Step 5: Tape temperature compensation and pulse heating:

[0129] Step 5-1: Infrared thermal imaging scanning: The equipment is configured as a FLIR A700 infrared thermal imager with a frame rate of 100 Hz, a temperature resolution of 0.1°C, and a scanning area of ​​5 mm × 5 mm.

[0130] Step 5-2: Low temperature area determination: If the temperature matrix variance exceeds 5°C2, it is determined to be a local low temperature area.

[0131] Step 5-3: Directed Pulse Heating: Use a Wagrave 500W quartz heating tube with a response time of less than 10 milliseconds. PID-PWM control: Use a 1% duty cycle step size and adjust the power every 10 milliseconds until the temperature stabilizes within the set range (±2°C).

[0132] Step 5-4: If the reheating fails for three consecutive times, the shutdown command is triggered and the error code is uploaded to the MES system, prompting the system to check the heater or tape material abnormality.

[0133] Step 6: Multispectral defect detection and process optimization:

[0134] Step 6-1: Visible light detection of bonding gap: Identify continuous areas with a gap greater than 0.3 mm between the tape edges, with a positioning accuracy of ±0.1 mm.

[0135] Step 6-2: Near-infrared adhesive layer penetration analysis: Calculate the effective contact area of ​​the adhesive layer using the reflectivity in the 1450 nm band, with the threshold set at 85%.

[0136] Step 6-3: Feature pyramid fusion: Fuse the visible light and near-infrared feature maps, and filter out bubble artifacts through morphological opening operation (kernel size 3×3).

[0137] Step 6-4: Gradient descent optimization: Optimize the tension, speed, and temperature parameter combinations by combining gap error, permeability, and bubble count. Use a learning rate of 0.001 and iterate until the loss function converges.

[0138] Step 6-5: Parameter storage and verification: The optimized parameters are encrypted and written to non-volatile memory for recall in the next cycle. If the new parameter verification fails, it automatically switches to the historical optimal parameter set and triggers the self-learning module.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent carton sealing control method based on multi-dimensional adaptive adjustment, characterized in that: The following steps are involved: Step 1: First, guide the carton to the carton sealing machine's infeed station. When the carton triggers the infeed sensor, an industrial camera simultaneously captures the carton's RGB image and depth data. The depth data includes a depth map, 3D point cloud data, normal vector data, confidence map, and pose parameters. The carton's geometric boundaries and pose parameters are extracted using a 3D reconstruction algorithm. Step 2: The visual data is fed into a multimodal feature fusion algorithm, which combines the carton surface texture features with the 3D topology to generate adaptive adjustment parameters for adjusting the width of the carton conveyor channel and the height of the tape pressing mechanism. Step 3: The conveyor belt guide plate spacing is driven according to the adaptive adjustment parameters, the width of the carton conveying channel is adjusted, and the height of the tape pressing mechanism is adjusted to seal the carton. At the same time, the tape rewind motor is dynamically controlled based on the tape deformation feedback to adjust the tape tension. Step 4: When the labeling sensor is triggered, the conveyor belt stops and a cutting path is generated through sub-pixel edge detection. The cutter mechanism is driven to cut the tape. At the same time, the tape is pressed again by the tape pressing mechanism to strengthen the tape bonding strength and complete the box sealing. Step 5: At the moment of cutting the tape, the temperature of the tape end is monitored in real time by an infrared thermal imager. If the temperature is lower than the viscosity activation threshold, the auxiliary heater is activated to pulse heat the cutting point. Step 6: After the carton is sealed, the conveyor belt is restarted to output the carton. Sealing defects are identified based on multispectral imaging and defect detection algorithms, and process optimization parameters are generated and iterated to the next work cycle.

2. The intelligent carton sealing control method based on multi-dimensional adaptive adjustment according to claim 1, characterized in that: Extracting the carton geometric boundary and posture parameters based on the 3D reconstruction algorithm in step 1 includes: Step 1-1: Perform bilateral filtering preprocessing on the depth map collected by the industrial camera to retain edge information while suppressing noise; Step 1-2: Use the random sampling consistency algorithm to fit the plane equations of the top and side surfaces of the carton and remove outliers; Step 1-3: Calculate the plane intersection line as the preliminary geometric boundary; Steps 1-4: Fusing normal vector data to optimize boundary curvature continuity; Steps 1-5: Perform weighted correction on the boundary points according to the confidence map to generate the final geometric model.

3. The intelligent carton sealing control method based on multi-dimensional adaptive adjustment according to claim 1, characterized in that: In step 2, the visual data is input into a multimodal feature fusion algorithm, and the adaptive adjustment parameters for adjusting the width of the carton conveying channel and the height of the tape pressing mechanism are generated by combining the carton surface texture features and the three-dimensional topological structure. The parameters include: Step 2-1: Perform gamma correction and histogram equalization preprocessing on the carton RGB image; Step 2-2: Extract local texture features of scratches, creases, and printed patterns on the carton surface through dilated convolution; Step 2-3: Construct a topological structure based on the 3D point cloud, and aggregate the curvature and normal vector differences of adjacent points; Step 2-4: Perform channel attention weighting on texture features and topological features; Step 2-5: Fusion of texture features and topological features through a gating mechanism to output adaptive adjustment parameters; Step 2-6: Perform L2 regularization on the fusion parameters to prevent overfitting.

4. The intelligent carton sealing control method based on multi-dimensional adaptive adjustment according to claim 1, characterized in that: The control method of the tape rewinding motor based on the tape deformation feedback in step 3 includes: Step 3-1: Deploy a high-speed polarization camera above the tape bonding area to capture the light intensity distribution on the tape surface at 1200 fps; Step 3-2: Calculate the local strain energy density based on the light intensity gradient change; Step 3-3: When the strain energy density exceeds the material yield threshold, an audible and visual alarm is triggered; Step 3-4: Calculate the tension compensation value through the fuzzy controller; Step 3-5: Dynamically adjust the speed of the tape rewinding motor using the PID algorithm; Step 3-6: Update the control command every 50ms to form a closed-loop feedback.

5. The intelligent carton sealing control method based on multi-dimensional adaptive adjustment according to claim 1, characterized in that: Generating a cutting path by sub-pixel edge detection in step 4 includes: Step 4-1: Construct a Gaussian pyramid in the trigger area of ​​the labeling sensor and perform 3-level downsampling; Step 4-2: Use the Sobel operator to extract the rough positioning area of ​​the tape edge; Step 4-3: Calculate the sub-pixel edge offset based on the Zernike moment; Step 4-4: Generate a smooth cutting path by fitting a cubic Bezier curve; Step 4-5: Use the inverse kinematics model to convert the path into servo motor pulse instructions; Step 4-6: Start the pressing roller 0.1 second before cutting and apply a dynamic pressure of 0.8-1.2 MPa.

6. The intelligent carton sealing control method based on multi-dimensional adaptive adjustment according to claim 1, characterized in that: The step 5 of starting the auxiliary heater to pulse heat the cutting point includes the following sub-steps: Step 5-1: Use an infrared thermal imager to obtain the temperature matrix of a 5mm×5mm area at the cutting point; Step 5-2: Calculate the variance of the temperature gradient distribution. If the variance exceeds the threshold, it is determined to be a local low temperature. Step 5-3: Start the auxiliary heater to compensate for the low temperature area; Step 5-4: Use PID-PWM controller to adjust the heating power; Step 5-5: Collect temperature feedback every 10ms until the temperature stabilizes within the set range; Step 5-6: If compensation fails three times in a row, a shutdown and maintenance instruction is triggered.

7. The intelligent carton sealing control method based on multi-dimensional adaptive adjustment according to claim 1, characterized in that: In step 6, identifying the sealing defects based on multispectral imaging and defect detection algorithms and generating process optimization parameters to iterate to the next working cycle include: Step 6-1: Detect the gap between the edges of the tape using visible light and mark the continuous area exceeding 0.3 mm. Step 6-2: Analyze the uniformity of adhesive layer penetration in the near-infrared band and calculate the effective contact area ratio; Step 6-3: Fuse multispectral feature maps through feature pyramid network; Step 6-4: Perform morphological opening operation filtering on the bubble defect to eliminate artifact interference; Step 6-5: Generate a heat map based on defect type and location; Step 6-6: Optimize the tension, speed, and temperature parameter combinations using the gradient descent algorithm; Step 6-7: Encrypt the optimized parameters and write them into the process parameter queue of the non-volatile memory; Steps 6-8: When the next box sealing cycle starts, extract the latest parameters from the queue and verify their validity; Step 6-9: If the parameter verification passes, the set values ​​of tape tension, conveying speed and heating temperature are dynamically updated; Step 6-10: If the parameter verification fails, the historical parameter set is enabled and the parameter self-learning module is triggered.

8. The intelligent carton sealing control method based on multi-dimensional adaptive adjustment according to claim 3, characterized in that: The gating mechanism in steps 2-5 includes: Step 2-5-1: Perform spatial attention weighting on the texture feature map to highlight the scratch area; Step 2-5-2: Perform channel attention weighting on the topological feature map to strengthen the curvature mutation points; Step 2-5-3: Modeling feature temporal dependencies through gated recurrent units; Step 2-5-4: Perform batch normalization on the fused features.

9. The intelligent carton sealing control method based on multi-dimensional adaptive adjustment according to claim 4, characterized in that: Calculating the tension force compensation value by the fuzzy controller in step 3-4 includes the following steps: Step 3-4-1: Define strain error and error change rate as fuzzy input variables; Step 3-4-2: Build a fuzzy rule base to cover all working condition combinations; Step 3-4-3: Use the centroid method to defuzzify and output the compensation value; Step 3-4-4: Perform dead zone compensation on the output value to prevent high-frequency oscillation.

10. The intelligent carton sealing control method based on multi-dimensional adaptive adjustment according to claim 5, characterized in that: Generating a smooth cutting path by cubic Bezier curve fitting in step 4-4 includes: Step 4-4-1: Select path control points at equal intervals along the tape edge based on the sub-pixel positioning results, and record the coordinates and normal vector directions of each point; Step 4-4-2: Construct a curvature continuity constraint equation to ensure that the curvature change rate between adjacent control points does not exceed 5%; Step 4-4-3: Use the particle swarm optimization algorithm to iteratively calculate the spatial coordinates of the control points. The objective function is to minimize the total curvature change rate of the path. Step 4-4-4: Perform quintic spline interpolation on the optimized control points to generate a cubic Bezier curve path with continuous curvature; Step 4-4-5: Discretize the curve path into 1000 interpolation points and convert them into the pulse sequence of the servo motor through the inverse kinematics model; Step 4-4-6: Verify the path smoothness in the simulation environment. If the maximum acceleration exceeds 1.5g, return to step 4-4-3 and re-optimize.

Images