Health care product package integrity detection method based on synergistic efficiency and intelligent detection

Abstract

The invention provides a health care product package integrity detection method based on cooperative efficiency and intelligent detection. The method comprises the following steps: S1, detecting time-space synchronization and hardware module calibration of a system; s2, collecting whole-domain space-time synchronous data of health care product packaging; s3, packaging sample feature preliminary screening and suspected sample extraction; s4, performing spatial visual three-dimensional point cloud accurate reconstruction on the suspected sample; s5, carrying out fusion mapping on the multi-modal sensor data and the three-dimensional point cloud model; s6, defect feature accurate identification and feature gradient feedback; s7, defect grade judgment and full-dimension cross check are carried out; s8, performing incremental training and full-process algorithm optimization on the detection model; and S9, carrying out production line cooperative execution and full-process feedback of a detection result. Through multi-algorithm two-way closed-loop fusion and full-process cooperation, the accuracy and reliability of package defect identification are improved, the real-time detection requirements of an industrial production line are met, meanwhile, full tracing of detection data and continuous optimization of detection performance are achieved, and the health care product package quality is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of health product packaging testing technology, and in particular to a method for testing the integrity of health product packaging that combines collaborative efficiency and intelligent testing. Background Technology

[0002] In the current health supplement packaging integrity testing industry, traditional testing methods mainly rely on manual visual inspection, which suffers from low testing efficiency, strong subjectivity, and high rates of missed and false detections, failing to meet the high-speed testing requirements of modern production lines. Existing automated testing solutions mostly employ single-sensor detection or simple data fusion technology, which has several limitations: First, the accuracy of 3D point cloud reconstruction is insufficient, making it difficult to capture subtle defects on the irregular curved surfaces of health supplement packaging, easily leading to reconstruction distortion; second, the fusion of multimodal sensor data and spatial vision data is not in-depth, resulting in severe feature loss and failing to fully reflect the physical characteristics and spatial features of the packaging; third, defect recognition algorithms mostly rely on traditional convolutional neural networks, making it difficult to retain the spatial structural information of defects, leading to missed detection of minor defects; fourth, the testing process is a one-way data transmission without a closed-loop optimization mechanism, preventing the testing accuracy and efficiency from continuously improving with the number of tests; fifth, the coordination between testing results and production line execution is poor, failing to achieve integrated control of testing-execution-optimization, and making it difficult to meet the stringent packaging quality standards of the health supplement industry. Summary of the Invention

[0003] This invention provides a collaborative and intelligent detection method for the integrity of health product packaging. Addressing the aforementioned shortcomings of existing technologies, it achieves automated, high-precision, and intelligent control of the entire health product packaging inspection process. It solves problems such as low efficiency of manual inspection, insufficient accuracy of automated inspection, feature loss, lack of closed-loop optimization, and poor production line coordination. Through multi-algorithm bidirectional closed-loop fusion and full-process collaboration, it improves the accuracy and reliability of packaging defect identification, adapts to the real-time inspection needs of industrial production lines, and simultaneously achieves full traceability of inspection data and continuous optimization of inspection performance, ensuring the quality of health product packaging.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: A collaborative efficiency and intelligent detection method for the integrity testing of health supplement packaging includes the following steps: S1: Perform spatiotemporal synchronization and accuracy calibration on the hardware module of the detection system to obtain the calibration parameter set of the detection system; S2: Based on the calibration parameter set of the detection system, perform full-domain spatiotemporal synchronous data acquisition on health product packaging samples and bind the sample's unique identification code to obtain the full-domain raw dataset; S3: An improved lightweight convolutional neural network algorithm is used to extract features and perform initial screening of samples from the original dataset across the entire domain to obtain a suspected sample dataset; S4: The radial basis function implicit surface reconstruction algorithm is used to preprocess the point cloud and accurately reconstruct the 3D point cloud of the suspected sample dataset to obtain a set of accurate spatial vision models. S5: The tensor product kernel spatial fusion algorithm is used to perform deep multimodal data fusion by combining the sensor feature set extracted from the suspected sample dataset with the spatial feature set extracted from the spatial vision precision model set to obtain the fused dataset; S6: The capsule network algorithm with graph attention enhancement is used to accurately identify defect features in the fused dataset, resulting in a defect identification dataset; S7: Based on the defect identification dataset, perform defect level determination and full-dimensional cross-validation to obtain the defect level determination and verification dataset; S8: Based on the defect level judgment and verification dataset, an incremental training sample set is constructed to perform incremental training and parameter optimization on the improved lightweight convolutional neural network algorithm, radial basis function implicit surface reconstruction algorithm, tensor product kernel space fusion algorithm and graph attention enhanced capsule network algorithm to obtain the model optimization parameter set; S9: Based on the defect level judgment and verification dataset and the model optimization parameter set, the detection system and the production line control system are coordinated to execute, and the detection results of the whole process are fed back to each step from S1 to S8 to form a closed loop of the whole process of health product packaging integrity detection.

[0005] In this specification, S5 generates spatial registration error during the deep fusion of multimodal data. This spatial registration error is fed back to S4 to adjust the parameters of the radial basis function implicit surface reconstruction algorithm. S6 generates defect feature gradient during the accurate identification of defect features. This defect feature gradient is fed back to S5 to adjust the parameters of the tensor product kernel spatial fusion algorithm.

[0006] In this specification, the specific process by which S5 feeds back the spatial registration error to S4 is as follows: S5 calculates the spatial registration error of the two types of features based on the sensor feature set and the spatial feature set. This spatial registration error is transmitted to the radial basis function implicit surface reconstruction algorithm of S4 in real time. Based on this spatial registration error, S4 adjusts the basis function weight coefficients and support radius of its own algorithm. The adjusted algorithm then performs accurate 3D point cloud reconstruction on the suspected sample dataset, generates an optimized spatial visual accurate model set, and inputs it into S5 to achieve dynamic optimization of reconstruction accuracy.

[0007] In this specification, the specific process by which S6 feeds back the defect feature gradient to S5 is as follows: When S6 performs defect feature recognition, it generates a set of digital capsule vectors through the capsule layer operation of the graph attention-enhanced capsule network algorithm. At the same time, it calculates the first-order partial derivative of the total model loss with respect to the set of digital capsule vectors to obtain the defect feature gradient. This defect feature gradient is transmitted to the tensor product kernel spatial fusion algorithm of S5 in real time. Based on this defect feature gradient, S5 adjusts the kernel weight matrix of its own algorithm. The adjusted algorithm recombines the sensor feature set and the spatial feature set to perform deep fusion of multimodal data, generates an optimized fusion dataset, and inputs it into S6 to improve the defect recognition accuracy of the fusion features.

[0008] In this specification, S4, S5, and S6 form a two-way interactive closed loop. The termination condition of the closed loop is: when the spatial registration error generated by S5 drops to the preset matching threshold, and the defect identification dataset output by S6 is verified as valid by the full-dimensional cross-validation of S7, the two-way interactive closed loop terminates. S4, S5, and S6 respectively output the final spatial visual precision model set, fusion dataset, and defect identification dataset, which are used for defect level determination, algorithm incremental training, and production line collaborative execution.

[0009] In this specification, the specific execution process of the improved lightweight convolutional neural network algorithm in S3 is as follows: spatial visual features and sensor physical features are extracted from the original dataset. The two types of features are concatenated to form a lightweight feature vector. This lightweight feature vector is input into the trained improved lightweight convolutional neural network model. The model outputs the category prediction probability of the sample through feature operation. Based on the preset probability threshold, samples that cannot be clearly determined as qualified or clearly defective are selected. All relevant data of such samples are integrated to form a suspected sample dataset.

[0010] In this specification, the specific execution process of the implicit surface reconstruction algorithm of radial basis function in S4 is as follows: 3D coarse-scan point cloud data is extracted from the suspected sample dataset. First, the 3D coarse-scan point cloud data undergoes point cloud preprocessing operations, including outlier removal and extraction of key points in potential defect areas. Then, based on the preprocessed point cloud data, a compactly supported radial basis function is constructed and the implicit surface model parameters are solved. Subsequently, a laser 3D scanner is controlled to perform high-precision 3D fine scanning on the suspected samples to obtain fine-scan point cloud data. The fine-scan point cloud data is fitted to the implicit surface model to complete the 3D point cloud reconstruction. Finally, the six-view visual features of the suspected samples are combined with the reconstructed point cloud model to bind spatial features and generate a precise spatial visual model set.

[0011] In this specification, the full-dimensional cross-validation in S7 includes three independent validation levels: sensor data validation, spatial vision data validation, and fusion feature validation. Sensor data validation verifies the physical matching of defect features based on the sensor feature set extracted from the suspected sample dataset. Spatial vision data validation verifies the spatial matching of defect features based on the spatial vision precision model set. Fusion feature validation verifies the fusion matching of defect features based on the fusion dataset. When all three validation levels pass, the defect identification dataset is deemed valid, and a corresponding defect level judgment and validation dataset is generated based on this defect identification dataset. When any validation level fails, the relevant data of the corresponding sample is re-incorporated into the suspected sample dataset, and a secondary detection process is initiated.

[0012] In this specification, the construction and incremental training process of the incremental training sample set in S8 is as follows: Select samples with valid judgment results from the defect level judgment and verification dataset, remove suspected false detection samples, classify and integrate the relevant data of the valid samples according to algorithm type, construct and improve the incremental training sample sets corresponding to the lightweight convolutional neural network algorithm, the radial basis function implicit surface reconstruction algorithm, the tensor product kernel space fusion algorithm, and the graph attention-enhanced capsule network algorithm, perform small-batch incremental training for the four types of algorithms and adjust the core parameters of each algorithm, and integrate the optimization parameters of all algorithms after training to form the model optimization parameter set.

[0013] In this specification, the specific process of collaborative execution between the detection system and the production line control system in S9 is as follows: Based on the defect level and detection effectiveness in the defect level judgment and verification dataset, corresponding processing suggestions are generated. Based on the processing suggestions, execution instructions corresponding to each actuator on the production line are generated. The execution instructions and the model optimization parameter set are transmitted to the production line control system. The production line control system controls the actuators to complete the corresponding actions of sample release, sorting and re-inspection, direct rejection or manual re-inspection, so as to realize the collaboration between detection and production.

[0014] In summary, the present invention has at least the following beneficial effects: It solves the problems of distorted 3D reconstruction and insufficient multimodal data fusion in existing automated inspection solutions. Through precise reconstruction and deep fusion technology, it fully preserves the spatial and physical characteristics of the packaging, significantly improves the accuracy of defect identification, and effectively avoids the missed detection and false detection of minor defects.

[0015] It achieves bidirectional closed-loop optimization of the testing process. Through parameter feedback and incremental training among the core algorithms, the performance of the testing system continuously improves with the number of tests. The algorithm model can be adaptively optimized without manual intervention, adapting to the testing needs of different types of health product packaging.

[0016] It achieves seamless collaboration between testing and production lines, directly translating test results into production line execution instructions. This enables precise sample sorting, release, rejection, and re-inspection, significantly reducing manual intervention and improving production line testing efficiency and quality control.

[0017] It enables full lifecycle traceability of testing data, constructs a complete testing archive covering sample production information, collected data, testing results and execution records, and provides reliable data support for subsequent quality analysis and process optimization.

[0018] It improves the stability and applicability of the testing system. Through precise hardware calibration and real-time monitoring of system status, it reduces the failure rate of the testing system. It can be adapted to the real-time testing of various health product packaging such as bottles, cans, bags, and boxes, and has strong versatility. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart of the health product packaging integrity detection method involving collaborative efficiency and intelligent detection involved in this invention.

[0020] Figure 2 This is a schematic diagram of the bidirectional feedback optimization closed loop of the three core algorithms S4-S6 involved in this invention.

[0021] Figure 3 This is a flowchart illustrating the incremental training process of the four core algorithms in steps S3-S6 involved in this invention.

[0022] Figure 4 This is a schematic diagram of the production line collaborative execution and full-process feedback involved in this invention. Detailed Implementation

[0023] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] like Figure 1 As shown, this embodiment provides a method for detecting the integrity of health product packaging with collaborative efficiency and intelligent detection, including: S1: Detection system spatiotemporal synchronization and hardware module calibration The core objective of this step is to achieve spatiotemporal synchronization calibration and accuracy standardization of all hardware modules in the detection system, eliminate system errors in hardware acquisition, and provide a unified time and spatial reference. A multi-sensor spatiotemporal synchronization calibration algorithm and laser 3D calibration technology are selected. The former achieves unbiased synchronization in the time dimension through hardware clock synchronization and data timestamp binding, while the latter unifies the spatial coordinates of the spatial vision acquisition unit and the multimodal sensor module through a standard calibration board, avoiding subsequent data fusion deviations caused by spatiotemporal asynchrony.

[0025] The complete hardware module of the detection system includes a conveyor position encoder, a spatial vision acquisition unit, a multimodal sensor module, an edge computing module, and a data interaction gateway. All hardware is industrial-grade detection equipment, which meets the real-time acquisition and calculation requirements of the health product packaging production line.

[0026] Core implementation process: 1. Hardware Module Physical Positioning: All hardware modules are fixedly installed according to the production line layout. The conveyor position encoder is installed at the conveyor drive shaft to collect the spatial position of the packaged samples. The spatial vision acquisition unit is installed above and around the inspection area, including a laser 3D scanner and a six-view industrial area scan camera. The multimodal sensor module is installed on both sides of the inspection area, completely overlapping with the acquisition range of the spatial vision acquisition unit. The edge computing module and data interaction gateway are deployed in the side control cabinet of the inspection area to realize real-time data transmission and processing.

[0027] 2. Time Synchronization Calibration: A unified clock synchronization command is issued to all hardware modules through the edge computing module to calibrate the local clock of each module with the high-precision clock of the edge computing module, and the calibration error is controlled within 1ms; a unified format timestamp is added to the data collected by all hardware modules, which includes year, month, day, hour, minute, second and millisecond, to ensure that the data collected by different modules can be accurately matched based on the timestamp and eliminate the time dimension acquisition deviation.

[0028] 3. Spatial Coordinate Calibration: A high-precision 3D standard calibration plate is used as the calibration reference. The spatial coordinate accuracy of the calibration plate is 0.001mm. The calibration plate is placed at the initial position in the detection area. The visual spatial coordinates of the calibration plate are obtained by scanning the calibration plate through a spatial vision acquisition unit. The sensor detection coordinates of the calibration plate are acquired through a multimodal sensor module. A linear mapping relationship between the visual spatial coordinates and the sensor detection coordinates is established, and its expression is: ; In the formula, This is the sensor detection coordinate vector, which includes the detection point coordinates of pressure, airtightness, laser displacement, and leakage sensors; It is a visual space coordinate vector, containing the coordinates of the acquisition points of the laser 3D scanner and the industrial area scan camera; The spatial calibration weight matrix is ​​obtained by fitting multiple sets of matched coordinates from the calibration plate. This is a spatial calibration offset term used to eliminate systematic errors in coordinate mapping. Through this mapping relationship, the spatial coordinates of the data acquired by all hardware modules are unified. The unified spatial coordinate system is a right-handed Cartesian coordinate system, with the origin being the initial spatial position of the detection area, and the spatial coordinate accuracy is controlled within 0.01mm.

[0029] 4. Hardware Accuracy Calibration: The spatial vision acquisition unit and the multimodal sensor module were calibrated separately. The laser 3D scanner verified the point cloud acquisition accuracy by scanning a standard calibration board, and the scanning parameters were adjusted to achieve a coarse scan accuracy of 0.1 mm and a fine scan accuracy of 0.01 mm. The industrial area array camera adjusted the focal length and exposure parameters by shooting a standard resolution board, so that the resolution of the six-view images all reached 4K, without blurring, overexposure, or underexposure. The multimodal sensor module adjusted the detection parameters by collecting standard reference values, and the accuracy of the pressure sensor reached 0.01 kPa, the airtightness sensor reached 0.01 mL / min, the laser displacement sensor reached 0.01 mm, and the detection sensitivity of the leakage sensor reached 0.01 mL.

[0030] This step outputs the calibration parameter set for the detection system. ,in For time synchronization parameters, This parameter set represents the optimal acquisition parameters for each hardware module. It is stored in the edge computing module and serves as the core basis for global data acquisition. All hardware modules acquire data based on this parameter set.

[0031] S2: Real-time and spatiotemporal synchronous data collection for health product packaging The core objective of this step is to achieve full-domain, spatiotemporal synchronous, and multi-dimensional data acquisition of health product packaging samples on the production line based on the calibration parameter set of step S1, obtaining complete raw data of the samples and providing a data foundation. Triggered synchronous acquisition technology is selected, using the position signal from the conveyor position encoder to trigger the synchronous operation of all acquisition units, ensuring that all data for the same sample are acquired in the same spatiotemporal dimension, avoiding acquisition deviations caused by sample movement. The acquired data covers spatial visual data and multimodal sensor physical data, achieving full-dimensional coverage of sample characteristics.

[0032] The core input for this step is the set of calibration parameters for the detection system output from step S1. The edge computing module calls this parameter set to configure the parameters of all hardware modules, and all acquisition units are configured according to this parameter set. Working with the optimal acquisition parameters, according to To achieve time synchronization, by and Achieve spatial coordinate unification.

[0033] Core implementation process: 1. Sample Transport and Position Triggering: Health product packaging samples are transported at a constant speed to the testing area via a production line conveyor. The conveying speed is matched to the production line cycle time and is set to 0.5 m / s. The conveyor position encoder collects the spatial coordinates of the sample in real time. When the geometric center of the sample reaches the initial spatial position of the testing area... At that time, the position encoder sends a position trigger signal to the edge computing module.

[0034] 2. Global synchronous acquisition trigger: After receiving the position trigger signal, the edge computing module immediately sends a synchronous acquisition command to the spatial vision acquisition unit and the multimodal sensor module. All acquisition units respond and start acquisition within 1ms, ensuring that all data acquisition of the same sample is carried out in the same spatiotemporal dimension, eliminating acquisition deviation caused by sample movement.

[0035] 3. Spatial visual data acquisition: The spatial visual acquisition unit is arranged according to... The parameters in the settings are used to start data acquisition. The laser 3D scanner performs a global 3D coarse scan of the sample, obtaining the 3D coarse scan point cloud data of the sample. The acquisition range covers the entire surface of the sample, with sampling points evenly distributed, encompassing all spatial contour features of the sample; a six-view industrial area array camera simultaneously captures images of the sample, resulting in a high-resolution six-view image set. The six perspectives are frontal, rear, left, right, top, and bottom, ensuring that the image covers the entire surface of the sample without any blind spots.

[0036] 4. Multimodal Sensor Data Acquisition: The multimodal sensor module and the spatial vision acquisition unit start synchronously to acquire the physical features of the sample across the entire area. The acquisition points correspond one-to-one with the points acquired by the spatial vision, resulting in raw detection data from four types of sensors: pressure sensor data. The contact pressure values ​​at various points on the sample surface are collected to reflect the hardness and integrity of the sample surface; airtightness sensor data. The amount of gas leakage at the sample seal is collected to reflect the sample's sealing performance; laser displacement sensor data... The distance deviation between the sensor and various points on the outer surface of the sample is collected to reflect the deformation of the sample surface; leakage sensor data. The leakage electrical signal value of the sample at the leakage-prone parts is collected to reflect the leakage prevention performance of the sample.

[0037] 5. Data Preprocessing and Integration: Each acquisition unit transmits the acquired data to the edge computing module in real time. The edge computing module first preprocesses the raw data, removing invalid data caused by hardware interference and interpolating missing data. Then, it adds a unified timestamp and spatial coordinates to all preprocessed data, based on step S1. and The spatial coordinates of all data are unified; finally, all data of the same sample are bound by the sample unique identification code ID, which is generated by the production line coding system and contains information such as the sample's production batch, product model, and production time, to ensure data traceability.

[0038] This step outputs the full-domain raw dataset of health supplement packaging. ,in To standardize timestamps, The dataset consists of unified spatial coordinates; it contains the complete raw data of each health product packaging sample in the production line and is transmitted to the edge computing module in real time.

[0039] S3: Initial screening of packaging sample characteristics and extraction of suspected samples The core objective of this step is to perform lightweight and rapid feature extraction and binary classification screening of health product packaging samples based on the full-domain raw dataset from step S2. This quickly distinguishes between qualified samples, suspected samples, and samples with clear defects, filtering out suspected samples for subsequent precise reconstruction and identification. This reduces the computational load of subsequent core algorithms and improves the overall detection efficiency. An improved lightweight convolutional neural network algorithm is selected. This algorithm simplifies the network structure based on traditional lightweight convolutional neural networks, retaining only the feature extraction layer and the binary classification output layer. It balances the effectiveness of feature extraction with lightweight computation, completing the initial screening of a single sample within 100ms. This adapts to the real-time detection requirements of the production line and avoids production line delays caused by complex calculations.

[0040] The core input for this step is the original dataset of the entire health product packaging domain, output from step S2. The edge computing module extracts features from all data in the dataset.

[0041] Core implementation process: 1. Lightweight global feature extraction: The edge computing module... Lightweight feature extraction is performed on the data, extracting spatial visual features and sensor physical features separately, and then concatenating the two types of features to obtain the global lightweight feature vector of the sample. Among them, spatial visual feature extraction targets and To carry out, on Extract geometric features from the point cloud, including 16-dimensional features such as point cloud quantity, point cloud centroid, and surface roughness; Color and texture features were extracted from the six-view images, with 8-dimensional features extracted from each image, totaling 48 dimensions across the six views and 64 dimensions of spatial visual features; sensor physical feature extraction was conducted targeting... The process involves extracting 10-dimensional features, including maximum, minimum, mean, variance, and peak value, from the data of each sensor type, resulting in a total of 40-dimensional features across the four sensor types. The 64-dimensional spatial visual features are then concatenated with the 40-dimensional sensor physical features to obtain a 104-dimensional global lightweight feature vector. This vector comprehensively reflects the core surface and physical features of the sample, and has low dimensionality, making it suitable for fast computation.

[0042] 2. Improved Lightweight Convolutional Neural Network Model Construction and Training: An improved lightweight convolutional neural network model is constructed, consisting of only three layers: an input layer, a feature extraction layer, and a binary classification output layer. The input layer has a 104-dimensional input dimension, along with a global lightweight feature vector. The dimensions are consistent; the feature extraction layer contains two convolutional layers and two pooling layers. The first convolutional layer has 32 3×1 convolutional kernels, and the second convolutional layer has 64 3×1 convolutional kernels. Both pooling layers use max pooling with a stride of 2, used for feature reduction and enhancement. The binary classification output layer is a fully connected layer with an output dimension of 3, corresponding to three categories of samples: qualified, suspected, and clearly defective. The softmax function maps the output values ​​to the [0,1] interval to obtain the probability of a sample belonging to each category. The model's loss function uses the cross-entropy loss function, whose expression is: ; In the formula, This represents the total number of samples used for model training. For sample category identification, This represents a qualified sample. Representing suspected samples, This represents a clearly defined defective sample; For the first The true class label of each sample is represented using one-hot encoding. If a sample belongs to class... but ,otherwise ; For the first Each sample belongs to category The predicted probability is obtained by the softmax function of the binary classification output layer of the model. The training samples for the model come from a historical dataset of health product packaging inspection, covering all common packaging types and defect types. The training convergence condition is set to the loss function value being less than 1. After training, the model parameters are stored in the edge computing module.

[0043] 3. Initial Sample Screening: The extracted global lightweight feature vectors will be used for... The improved lightweight convolutional neural network model, after training, is used as input. The model's forward propagation yields the predicted probabilities of samples belonging to three categories: qualified, suspected, and clearly defective. The edge computing module performs an initial screening based on the predicted probability, and the judgment rule is: if If it is, then it is judged as a qualified sample; if If the sample meets the above two conditions, it is considered a clearly defective sample; if neither of the above two conditions is met, regardless of... Regardless of the value, all samples are classified as suspected samples. This judgment rule balances the accuracy and rigor of the initial screening, classifying all samples that cannot be clearly determined as suspected samples to avoid missed detections.

[0044] 4. Initial screening result data division: Based on the initial screening judgment result, the edge computing module divides the global original data set in step S2 into three sub-data sets according to the sample unique identification code ID: qualified sample data set , which contains all the sample data judged to be qualified, and is directly transmitted to the production line execution module. The qualified products are released; the clear defect sample data set , which contains all the sample data judged to have clear defects, is transmitted to the production line execution module. The defective products are judged and removed. At the same time, this data set is stored in the historical sample library for subsequent incremental training of the model; the suspected sample data set , which contains all the sample data judged to be suspected, retains all dimensions and details of the original data, and is used as the input data for the accurate reconstruction of the visual three-dimensional point cloud of the suspected sample space.

[0045] This step outputs three types of data sets, namely , , , among which is the core output, retaining the complete original data of the suspected samples; and are directly used for the preliminary execution of the production line to achieve rapid sorting of samples and improve detection efficiency.

[0046] S4: Accurate reconstruction of the visual three-dimensional point cloud of the suspected sample space The core goal of this step is to achieve high-precision three-dimensional point cloud reconstruction of suspected samples based on the suspected sample data set in step S3, generate a continuous and smooth final three-dimensional solid point cloud model, and provide an accurate spatial feature base for subsequent multi-modal data fusion. The radial basis function implicit surface reconstruction algorithm is selected. This algorithm is different from the traditional Poisson reconstruction and greedy projection triangulation algorithms, and is suitable for high-precision reconstruction of small-batch rough scanned point clouds. It can accurately capture the irregular surface details of health product packaging through the local characteristics of the compactly supported radial basis function, and at the same time can receive the feedback spatial registration error to achieve dynamic parameter correction. It not only solves the problems of reconstruction distortion and noise residue of traditional algorithms, but also realizes two-way interactive optimization, which is the core link connecting spatial visual data and multi-modal sensor data.

[0047] The core input of this step is the suspected sample data set output in step S3 , and the core data for three-dimensional point cloud reconstruction is extracted from this data set: three-dimensional rough scanned point cloud data , spatial initial coordinates , and six-view high-definition image set acquisition trigger signal. All the extracted data are bound to the sample unique identification code ID.

[0048] 1. Preprocessing of rough scanned point cloud: For the three-dimensional rough scanned point cloud data Preprocessing is performed to eliminate outlier interference and extract key points, laying the foundation for subsequent basis function center selection. First, statistical filtering is used to remove outliers, and then calculations are performed. The mean and standard deviation of the distances between each point and its 50 neighboring points are calculated. Points whose mean distance exceeds three times the standard deviation are identified as outliers and removed, resulting in a denoised coarse point cloud. Then, the SIFT-3D algorithm is used to extract... The key points are extracted from the edges, corners, and potential defect areas of health product packaging. These potential defect areas include the bottle opening seal, can opening rolled edges, bag heat-sealed edges, box folded edges, easily damaged areas on the packaging surface, and label affixing areas, resulting in a key point set. The key point set preserves the core spatial features of the point cloud, reducing the amount of subsequent computation.

[0049] 2. Construction of the Radial Basis Function Implicit Surface Reconstruction Algorithm Model: The core of constructing the radial basis function implicit surface reconstruction algorithm model is to build implicit surface functions for the 3D point cloud, realizing the transformation from a coarsely scanned point cloud to a continuous and smooth 3D surface. First, the Wendland C2 compactly supported radial basis function is selected as the basis function, and its expression is: ; In the formula, The normalized radial distance is calculated as follows: , The Euclidean norm is used to calculate the straight-line distance between two points in space. Let be the coordinates of any point to be reconstructed in three-dimensional space; For the first The radial basis function centers are formed by the key point set. Select directly; The support radius of the basis function is dynamically adjusted according to the packaging type of health products. The support radius of bottle and can packaging is 5mm, and the support radius of bag and box packaging is 3mm. This is a truncation function, and its value is determined by the rule that when... hour, ,when hour, This is used to define the local scope of the basis function. Based on this basis function, the basic form of the implicit surface function is constructed: ; In the formula, The number of basis function centers is given by the keypoint set. The number of points is determined; For the first The weight coefficients of each basis function are the core parameters that the model needs to solve for; The coefficients of the polynomial terms, For constant terms, They are respectively The three-axis linear coefficients are used to correct the global trend of the surface and avoid surface distortion caused by excessive local features. This ensures that the reconstructed surface is rigorously processed from a denoised coarse-scanned point cloud. Add interpolation constraints: For any point in All satisfy To avoid singularities in the model solution, orthogonal constraints are added: , , , .

[0050] 3. Model parameter solution and initial surface reconstruction: Combining interpolation constraints and orthogonal constraints, a system of linear equations is constructed, with the unknowns of the system being... and The number of equations is , For the denoised coarse scan of point clouds The number of points is 4, and the number of orthogonal constraint equations is 4. The conjugate gradient method is used to solve the linear equation system to obtain the basis function weight coefficients. With polynomial coefficients The optimal solution is obtained; the optimal solution is substituted into the implicit surface function to obtain the initial three-dimensional implicit surface model. Based on this initial model, combined with the initial spatial coordinates The laser 3D scanner was controlled to perform a 3D fine scan on the suspected sample with a precision of 0.01 mm, resulting in 3D fine scan point cloud data. ,Will Fitting the model to the initial 3D implicit surface model yields a preliminary reconstructed 3D solid point cloud model. .

[0051] 4. Receive feedback and dynamically correct parameters: The initially reconstructed 3D solid point cloud model After transmission to subsequent steps, the spatial registration error is received as feedback. This error reflects the spatial coordinate deviation between the reconstructed model and the multimodal sensor data; the spatial registration error... Based on this, the parameters of the implicit surface reconstruction algorithm of radial basis functions are dynamically adjusted, and the adjustment formula is as follows: ; In the formula, The corrected algorithm parameter set contains ; The initial parameter set before correction; The learning rate of the algorithm is set to 0.001, balancing the effectiveness and stability of parameter correction. The gradient of the implicit surface function is calculated as follows: It reflects the changing trend of the surface at various points in space, and enables targeted correction of deviation areas.

[0052] 5. Final 3D solid point cloud model generation: This involves generating the corrected parameter set. Substituting the implicit surface function back into the model yields the final three-dimensional implicit surface model. ; 3D fine-scan point cloud data With a six-view high-definition image set The visual features are mapped to the final model, achieving the binding of spatial features with visual features. This ensures that each spatial coordinate point of the model corresponds to a clear visual feature, generating the final 3D solid point cloud model. .

[0053] This step outputs a set of precise spatial visual models. ,in The final 3D entity point cloud model is the core output of this step. It has the characteristics of being continuous, smooth, and highly accurate, and accurately reflects the full-dimensional spatial and visual features of the suspected sample. This dataset is bound to the sample's unique identification code ID, serving as the core spatial feature basis for the fusion mapping of multimodal sensor data and 3D point cloud model.

[0054] S5: Fusion mapping of multimodal sensor data and 3D point cloud model The core objective of this step is to deeply fuse the multimodal sensor physical data from step S3 with the 3D solid point cloud model from step S4, generating an integrated multimodal fusion detection model that includes both physical and spatial features. This provides a complete and comprehensive feature base for subsequent defect identification. A tensor product kernel space fusion algorithm is selected. This algorithm maps one-dimensional sensor physical data and high-dimensional spatial visual data to a unified tensor product kernel space, preserving the original features of both types of data to the maximum extent and avoiding the feature loss problem of traditional fusion algorithms. It can also receive feedback defect feature gradients to dynamically adjust kernel weights and calculate spatial registration errors, feeding them back to step S4 for reconstruction parameter correction. This is the core algorithm for achieving deep multimodal data fusion and bidirectional interactive optimization.

[0055] The input data for this step contains two core sources, both of which are bound to a unique sample identifier ID: one is the spatial vision precision model set output from step S4. Extract spatial feature sets from them First, it provides an accurate spatial benchmark for multimodal data fusion; second, it provides a dataset of suspected samples output from step S3. Extract sensor feature sets from them ,in The sensor data after spatial coordinate calibration in step S1 eliminates system errors.

[0056] Core implementation process: 1. Feature set preprocessing: Preprocessing the spatial feature set... With sensor feature set Preprocessing is performed to ensure data consistency and validity. This includes processing the spatial feature set. Feature vectorization is performed to transform the final 3D solid point cloud model. Spatial coordinates, surface curvature, visual grayscale values, and 3D fine-scanned point cloud data The point cloud coordinates, point cloud density, and other features are transformed into a fixed-dimensional spatial feature vector. All suspected samples had their spatial feature vectors unified to 1024 dimensions; for the sensor feature set Perform min-max normalization processing, and Map all data values ​​to the [0,1] interval to eliminate the dimensional differences between different sensor data, and obtain the normalized sensor feature vector. The dimensions are unified to 4 dimensions, corresponding to the normalized data of the four types of sensors.

[0057] 2. Tensor Product Kernel Space Fusion Algorithm Model Construction: The core of constructing the tensor product kernel space fusion algorithm model is to build a unified tensor product kernel space to achieve the organic fusion of sensor features and spatial features. First, regeneration kernel Hilbert spaces are constructed for the two types of features respectively. The regeneration kernel Hilbert space corresponding to the sensor features is... The regenerative kernel Hilbert space corresponding to the spatial characteristics is For both types of spaces, the Gaussian kernel function is used as the regeneration kernel function. The expression for the Gaussian kernel function of the sensor features is: ; In the formula, Let be any two feature vectors in the sensor feature set; The kernel width is the kernel space of the sensor kernel, which is a parameter that the model needs to solve for; Let L2 be the norm, used to calculate the squared distance between two eigenvectors. The Gaussian kernel function expression for spatial features is: ; In the formula, Let be any two feature vectors in the spatial feature set; The kernel width is the parameter that the model needs to solve for in the kernel space. A tensor product operation is performed on the two regenerated kernel Hilbert spaces to construct a unified tensor product kernel space. Any fused feature in this space can be represented as the tensor product of the sensor feature vector and the spatial feature vector, i.e. The rules for tensor product operations are as follows: If for dimensional vector, for If a vector is given by a given dimension, then its tensor product is: A dimensional vector, where each element is a pairwise product of the elements of the two original vectors, ensures comprehensive fusion of the two features. Based on this tensor kernel space, a multimodal fusion feature tensor is constructed, whose expression is: ; In the formula, This is the multimodal fusion feature tensor, which is the core feature after fusion; The tensor product kernel weight matrix is ​​the core solution parameter of the model, used to adjust the fusion weights of sensor features and spatial features; The kernel matrix of the sensor feature set has dimensions of . , The number of samplings by the sensor is given, and the matrix elements are the Gaussian kernel function values ​​of any two sensor eigenvectors. Let be the kernel matrix of the spatial feature set, with dimension . , For 3D fine-scan point clouds The number of points, where the matrix elements are the Gaussian kernel function values ​​of any two spatial eigenvectors; This is a tensor product operator used to achieve deep fusion of two kernel matrices.

[0058] 3. Model Parameter Solving and Initial Fusion Feature Tensor Generation: A model training sample set is constructed, with samples drawn from historical datasets of health product packaging inspections, covering all common packaging types and defect types. Each sample set is a paired sample of sensor features and spatial features, labeled as qualified / defective. The Adam optimizer is used to solve for the model parameters, with the core solution being the kernel width. With kernel weight matrix The model uses cross-entropy loss as its loss function, and the convergence condition is set to the loss function value being less than... Substitute the optimal parameters obtained from the solution into the multimodal fusion feature tensor formula, and then process the preprocessed sensor feature set. Spatial feature set Substituting the corresponding Gaussian kernel functions, we obtain the sensor kernel matrix and the spatial kernel matrix. After tensor product operation, these are multiplied by the kernel weight matrix to obtain the initial fused feature tensor. .

[0059] 4. Calculate the spatial registration error and feed it back to step S4: based on the initial fused feature tensor Calculate the spatial registration error between sensor features and spatial features. This error measures the coordinate deviation between the two types of features in the tensor kernel space. The smaller the deviation, the higher the fusion accuracy. Its expression is: ; In the formula, The dimension of the sensor feature vector. is the dimension of the spatial feature vector; For the first Each sensor feature vector; The linear projection function projects spatial feature vectors onto the physical dimension of the sensor, achieving coordinate matching between two types of features. Its expression is: , The projection weight matrix is... These are projection bias terms, all obtained by fitting from the training samples; The L2 norm is used to calculate the squared distance between the projected spatial feature vector and the sensor feature vector. The calculated spatial registration error is then used... The feedback is sent to step S4 in real time, serving as the core basis for parameter correction of the implicit surface reconstruction algorithm of the radial basis function in step S4.

[0060] 5. Receive feedback and dynamically adjust kernel weights: Adjust the initial fused feature tensor After being transmitted to subsequent steps, the defect feature gradient is received as feedback. This gradient reflects the changing trend of defect features; the larger the gradient, the more obvious the defect features in the corresponding region. (The text then abruptly shifts to a seemingly unrelated topic: defect feature gradient.) Based on this, the kernel weight matrix is... Dynamic adjustments are made, and the adjustment formula is as follows: ; In the formula, This is the adjusted final kernel weight matrix; The initial kernel weight matrix before adjustment; This is the learning rate for the algorithm, with a value of 0.001. The gradient of the fusion feature tensor reflects the changing trend of the tensor at various points in space. Through this adjustment, the fusion weight of the defective region is strengthened and the redundant features of the non-defective region are weakened, thereby improving the recognition of defects by the fusion features.

[0061] 6. Generation of multimodal fusion detection model: The adjusted final kernel weight matrix Substituting the formula for the multimodal fusion feature tensor back into the equation, we obtain the final fusion feature tensor. This tensor comprehensively integrates the physical features of the sensor with the spatial visual features of the 3D point cloud, and enhances the features of the defect region; the final fused feature tensor The final 3D solid point cloud model of step S4 Spatial feature binding is performed, mapping each feature value of the tensor to a specific location. Corresponding spatial coordinates This ensures that each spatial coordinate point corresponds to a complete fusion feature, generating a multimodal fusion detection model. .

[0062] This step outputs the fused dataset. ,in The multimodal fusion detection model is the core output of this step, realizing the integrated fusion of physical and spatial features, and providing an accurate and comprehensive feature foundation for subsequent defect identification. For the final fused feature tensor, For spatial registration error, The dataset uses unified spatial coordinates and is bound to a unique sample identification code (ID), serving as the sole input data for accurate identification of defect features.

[0063] S6: Precise Defect Feature Identification and Feature Gradient Feedback The core objective of this step is to achieve accurate identification and numerical output of defect type, size, and spatial location of suspected samples based on the multimodal fusion detection model of step S5. Simultaneously, it calculates defect feature gradients and feeds them back to step S5 to optimize fusion weights. A graph attention-enhanced capsule network algorithm is selected. This algorithm combines the advantages of graph attention mechanisms and capsule networks. Graph attention mechanisms strengthen the feature weights of potential defect regions, while capsule networks preserve the spatial structure information of defects, solving the problems of lost spatial features and missed detection of minor defects in traditional convolutional neural network algorithms. It also enables bidirectional interaction with step S5, making it the core algorithm for accurate defect identification. The closed-loop process of bidirectional feedback optimization of the three core algorithms S4-S6 is as follows: Figure 2 As shown.

[0064] The core input for this step is the fused dataset output from step S5. Extracting a multimodal fusion detection model from it. With the final fusion feature tensor As input to the algorithm, Provides a spatial reference for defect identification. It provides core fusion features for defect identification, and all input data are bound to the sample's unique identification code ID.

[0065] Core implementation process: 1. Graph attention feature extraction: The final fused feature tensor Construct as a graph structure ,in Let be the set of nodes in the graph. To define the dimension of the fusion feature tensor, each node corresponds to a feature dimension of the tensor, representing the fusion feature of a certain spatial point of the suspected sample. Let be the set of edges in the graph, where the edge weights represent the strength of the feature association between two nodes. Based on this graph structure, the attention weight of each node is calculated using a graph attention mechanism, expressed as: ; In the formula, For nodes For nodes Attention weights; This represents the weight vector for the attention mechanism; The graph attention linear transformation matrix is ​​used to map high-dimensional node features to a low-dimensional space. This is a vector concatenation operator; For nodes A set of 10 neighboring nodes is used to capture local feature associations; The activation function is set to a negative slope of 0.2 to achieve a non-linear mapping of features. Based on the attention weights, the graph attention-enhanced feature matrix is ​​obtained by weighted summation of features, and its expression is: ; In the formula, To enhance the feature matrix for graph attention; The ReLU activation function is used to enhance the nonlinear expression of features and remove invalid negative features. This matrix removes redundant nodes with weights below 0.01, retaining only the core features associated with defects, thus improving the efficiency of subsequent identification.

[0066] 2. Construction of Graph Attention Enhanced Capsule Network Algorithm Model: A graph attention enhanced capsule network algorithm model is constructed, consisting of four layers: a graph attention feature extraction layer, a capsule layer, a dynamic routing layer, and a defect classification and localization layer. The capsule layer is divided into primary capsules and digital capsules. The primary capsule transforms the graph attention enhanced features into 16-dimensional capsule vectors, preserving the spatial structure information of the features. Its expression is: ; In the formula, For the first One primary capsule vector; This is the primary capsule weight matrix; The first attention-enhanced feature matrix of the graph Row feature vectors. A total of 7 digital capsules are set, corresponding to 6 categories of health product packaging defects and 1 category of qualified samples. The 6 categories of defects are packaging damage, seal leakage, liquid leakage, label curling, label printing omission, and surface deformation. Each digital capsule corresponds to one category, and its capsule vector is the global defect feature. The dynamic routing layer updates the routing weights through three iterations to achieve accurate aggregation of primary capsule features into digital capsules. The routing weight update formula is: ; In the formula, Primary capsules With digital capsules Routing weights; This represents the number of iterations. For the first Primary capsules in the next iteration Mapping to digital capsules eigenvectors; For the first Digital Capsule after the next iteration The temporary output vector; This is the vector dot product operator. The normalized route weights are obtained after softmax normalization. The final expression for the digital capsule vector is: ; In the formula, For the first The final capsule vector of the number capsules; Given the sigmoid activation function, calculate the activation coefficients of the capsule vectors; This is the transformation matrix from the primary capsule to the digital capsule; Using the Euclidean norm, the vector magnitude is calculated, where the magnitude represents the degree of defect and the direction represents the type of defect. The defect classification and localization layer transforms the digital capsule vector into a numerical defect identification result. Defect type classification is achieved through a fully connected layer and a softmax function, spatial location of defects is achieved through a coordinate regression subnetwork, and defect size is calculated by combining the spatial features of the 3D fine-scanned point cloud.

[0067] 3. Precise Defect Feature Identification and Numerical Output: The final capsule vectors of the seven digital capsules are input into the defect classification and localization layer. First, the probability of a sample belonging to each category is obtained through the softmax function. The category corresponding to the maximum probability is taken as the defect type, and the defect type is numerically encoded as follows: qualified sample = 0, packaging damage = 1, seal leakage = 2, liquid leakage = 3, label curling = 4, label misprint = 5, surface deformation = 6. Then, based on the modulus of the digital capsule vectors, a multimodal fusion detection model is used. The spatial characteristics of the defect are analyzed to calculate its actual physical size, including the area of ​​planar defects and the volume of three-dimensional defects, resulting in a numerical representation of the defect size. Finally, the three-dimensional spatial coordinates of the defect center are output through a coordinate regression subnetwork. These coordinates are precisely mapped to the final 3D solid point cloud model. The system uses a spatial coordinate system and outputs the spatial boundary coordinates of the defect area to achieve precise spatial positioning of the defect. By integrating the defect type code, defect size, and defect spatial coordinates, a defect feature vector is obtained. .

[0068] 4. Calculate the defect feature gradient and feed it back to step S5: Based on the defect identification results and the total model loss, calculate the defect feature gradient. The gradient is the first-order partial derivative of the total model loss with respect to the final set of digital capsule vectors, and its expression is: In the formula, Let be the total loss of the model, and be the weighted sum of the classification loss and the localization loss; This is the final set of digital capsule vectors; the gradient reflects the changing trend of defect characteristics and is fed back to the S5 step in real time, serving as the core basis for the dynamic adjustment of the kernel weight matrix in the S5 step.

[0069] 5. Verification of identification results: Spatial location coordinates of the defect. Mapping to a multimodal fusion detection model Extract the fusion feature value corresponding to the coordinate and verify whether the feature value matches the defect type. For example, the airtightness sensor data corresponding to a sealing leak should significantly exceed the threshold. If they match, the identification result is deemed valid. If they do not match, the sample is re-included in the suspected sample dataset of step S3 for secondary detection.

[0070] This step outputs the defect identification dataset DS6={ , },in This is a defect feature vector, containing the defect type code and defect size; The three-dimensional spatial coordinates of the defect center; The dataset represents the gradient of defect features; the judgment result is qualified / specific defect type; this dataset is bound to the unique identification code ID of the sample and transmitted to subsequent steps, serving as the core basis for defect level judgment and cross-validation.

[0071] In some embodiments, the radial basis function implicit surface reconstruction algorithm of S4 and the graph attention-enhanced capsule network algorithm of S6 interact in pairs. The interaction is an indirect collaborative interaction, the core of which is the linkage between defect region localization and reconstruction focus optimization. The defect spatial localization result of S6 is fed back to S4, enabling S4 to optimize the 3D point cloud reconstruction accuracy of the defect region in a targeted manner, avoiding redundant calculations in global reconstruction, while improving the spatial feature integrity of the defect region, providing a more accurate spatial basis for subsequent fusion and recognition.

[0072] S6 employs a graph attention-enhanced capsule network algorithm to accurately identify defect features from the fused dataset, outputting a defect identification dataset containing the three-dimensional spatial coordinates of the defect centers. The system retrieves the boundary coordinates of the defect area; S6 feeds back these spatial coordinates to S4 in real time. Based on these coordinates, S4 locates the potential defect area reconstructed from the 3D point cloud and dynamically adjusts the parameters of the implicit surface reconstruction algorithm for the radial basis function in this area (focusing on optimizing the support radius of the basis function and the density of key point selection). This allows for local, precise reconstruction of the defect area, rather than global reconstruction. The optimized spatial visual precision model set is then input into S5, fused by S5, and input into S6, achieving targeted optimization of defect area reconstruction and recognition. This interaction, along with the interactions between S4-S5 and S5-S6, further improves the overall accuracy and efficiency of detection.

[0073] Based on the defect space coordinates fed back by S6, S4 locally adjusts the basis function support radius of the defect region. The adjustment formula is as follows: ; In the formula, The radius of the local basis function support for the defect region. Where k is the global support radius, and k is the adjustment coefficient. To reconstruct the coordinates of any point in space within the region, S6 outputs the three-dimensional spatial coordinates of the defect center; simultaneously, S4 calculates the reconstruction error of the defect region. This is used to verify the effect of local reconstruction; the formula is: In the formula, The number of finely scanned point clouds in the defect area. The actual fine-scan point cloud coordinates of the defect area. These are the coordinates of the point cloud after local reconstruction.

[0074] In some embodiments, the pairwise interactions of S4-S6 are a collaborative iterative process, not a single interaction ending independently. The overall interaction must simultaneously meet three conditions to end: 1. Spatial registration error output by S5 during the interaction between S4 and S5. The feature is reduced to within a preset feature matching threshold to ensure accurate matching between spatial features and sensor features; 2. In the interaction between S5 and S6, the defect identification dataset output by S6 was verified as valid by full-dimensional cross-validation in S7, and the total loss of the graph attention-enhanced capsule network algorithm was [missing information]. Reduce the loss to a preset threshold to ensure accurate defect identification results; 3. During the interaction between S4 and S6, the local reconstruction error of the defect area output by S4. Reduce the accuracy to within the preset reconstruction accuracy threshold to ensure the reconstruction accuracy of the defective area; When all three conditions are met, the pairwise interactions of S4-S6 terminate, and the final spatial vision precision model set, fusion dataset, and defect identification dataset are output respectively. If any condition is not met, the iterative interaction continues until all conditions are met. If the number of iterations reaches the preset limit and the conditions are still not met, an early warning is triggered, and the corresponding sample is assigned to manual re-inspection.

[0075] The quantitative condition for the end of the interaction is: ; In the formula, A threshold is preset for spatial registration error. To pre-determine the total loss of the capsule network algorithm with enhanced graph attention, A threshold is preset for the local reconstruction error of the defect area; where "the defect identification result is valid" is the qualitative result after S7 full-dimensional cross-validation, and must be satisfied at the same time as the above quantitative conditions before the interaction can be terminated.

[0076] S7: Defect Level Determination and Comprehensive Cross-validation The core objective of this step is to establish an industrial-grade defect grading standard based on the defect identification results of step S6, enabling the grading of defect samples. Simultaneously, it combines the raw and fused data from steps S3-S5 for comprehensive cross-validation to ensure the accuracy of defect identification and grading results, avoiding false positives and false negatives. Multi-feature fusion cross-validation technology is employed, achieving verification through multi-dimensional matching of defect features with raw sensor data and fused features. The defect grading is based on the packaging quality standards of the health supplement industry and the actual needs of the production line, balancing quality control and production efficiency.

[0077] The input data for this step contains two core sources, both bound to the sample's unique identifier ID: one is the defect identification dataset DS6={ output from step S6}. , The first is the core basis for determining the defect level; the second is the suspected sample dataset in step S3. S4 Step Spatial Visual Precision Model Set The fusion dataset in step S5 This serves as the foundational data for cross-validation across all dimensions, ensuring the comprehensiveness of the validation.

[0078] Core implementation process: 1. Industrial-Grade Defect Classification Standards: Based on national standards for packaging quality in the health product industry and the actual implementation needs of production lines, and combining three core dimensions—defect type, defect size, and defect location—a three-level defect classification standard has been established. The weights for the three dimensions are 0.4, 0.3, and 0.3, respectively, determined based on the degree of impact of each dimension on packaging quality. Level 1 Defects (Minor Defects): Defect types include non-functional defects such as label curling and missing label printing; defect size is less than a preset threshold (planar defect area < 5). 3D defect volume <3 Level 2 defects (moderate defects): Defects located in non-core areas of the packaging (such as non-sealed areas on the sides of the packaging); Level 2 defects (moderate defects): Defect types include surface deformation, minor packaging damage, and other semi-functional defects, with defect sizes within the preset threshold range (5). ≤ Planar defect area < 15 3 ≤3D defect volume<10 Level 3 defects (severe defects): defects are functional defects such as seal leakage, leakage, or severe packaging damage, and the defect size is greater than the preset threshold (planar defect area ≥ 15). 3D defect volume ≥10 (This refers to defects in the packaging itself, such as those located in the core area of ​​the packaging, like the bottle opening seal, the rolled edge of the can opening, or the heat-sealed edge of the bag.) Each dimension is quantitatively scored, with a total score of 100 points. Defects are classified into three levels based on the score: 80-100 points are Level 1 defects, 40-79 points are Level 2 defects, and 0-39 points are Level 3 defects.

[0079] 2. Defect Sample Level Quantification and Determination: Based on the defect level classification standard, samples identified as defects in step S6 are quantitatively scored and their levels determined. First, the defect identification dataset is extracted. The defect type, size, and location information are used to score each dimension according to preset rules: For defect type, functional defects score 0-30 points, semi-functional defects score 31-60 points, and non-functional defects score 61-100 points; for defect size, defects smaller than the threshold score 71-100 points, those within the threshold range score 31-70 points, and those larger than the threshold score 0-30 points; for defect location, non-core areas score 71-100 points, secondary core areas score 31-70 points, and core areas score 0-30 points. Then, the comprehensive score of the sample is calculated according to the dimension weights. The comprehensive score formula is: ; In the formula, The overall score for defects in the sample; Score the defect type dimension; Scoring is based on the defect size dimension; Scoring is given for the defect location dimension. Based on the overall score... Determine the defect level: It is a Level 1 defect. It is a level 2 defect. The defect is classified as Level 3; qualified samples are not graded and are directly marked as defect-free.

[0080] 3. Full-dimensional cross-validation: Combining the basic data from steps S3-S5, a full-dimensional cross-validation is performed on the defect identification results and the grade determination results to ensure the accuracy of the results. The validation is divided into three levels: First, sensor data validation, which verifies the spatial coordinates of the defect. Mapping to multimodal sensor raw data in step S3 First, extract the sensor detection value corresponding to the coordinates and verify whether the detection value exceeds the threshold range of qualified samples. For example, the leakage amount of the airtightness sensor corresponding to the sealing leakage defect should be significantly greater than the qualified threshold. If it does not exceed the threshold, it is judged as a suspected false detection. Second, spatial visual data verification is performed by mapping the spatial boundary coordinates of the defect area to the final three-dimensional solid point cloud model of step S4. With a six-view high-definition image set The process involves three steps: First, observing the spatial structure of the defect area using point cloud models to check for anomalies. Second, examining the visual features of the defect area using high-resolution images to see if they match the defect type. For example, packaging damage defects should show obvious signs of damage in the image; if these are not observed, it is considered a suspected false positive. Third, feature fusion verification is performed, combining the defect feature vectors... The final fused feature tensor of step S5 The matching process is performed to verify whether the feature values ​​of the defect region in the fused feature tensor are consistent with the features of the defect feature vector. If they are inconsistent, it is judged as a suspected false detection.

[0081] 4. Verification Result Processing: The results of the full-dimensional cross-validation are classified. If all three levels of verification pass, the defect identification result and the level determination result are deemed valid, and the final determination result is generated. If any level of verification fails, it is determined to be a suspected false detection, and all data of that sample are integrated and re-included in the suspected sample dataset of step S3. The secondary testing process will be initiated. The secondary testing will employ higher data collection precision and stricter judgment rules to ensure the accuracy of the test results. If the secondary testing still cannot make a clear judgment, the sample will be marked as a manual re-inspection sample and transferred to the manual re-inspection station on the production line for manual testing by professional quality inspectors.

[0082] This step outputs a defect level determination and verification dataset DS7={ID, defect type, defect size, defect location, defect level, determination result validity, and handling suggestions}. The handling suggestions are based on the defect level and the validity of the determination result: qualified samples are recommended to be released directly; Level 1 defect samples are recommended to be released after a warning; Level 2 defect samples are recommended to be re-inspected after sorting; Level 3 defect samples are recommended to be directly rejected; suspected false positive samples are recommended to be re-inspected; and manually re-inspected samples are recommended to be manually inspected offline. This dataset is bound to the unique sample identification code ID and is also transmitted to subsequent steps for incremental model training and production line collaborative execution, respectively.

[0083] S8: Incremental Training of Detection Model and Optimization of the Entire Algorithm Process The core objective of this step is to construct an incremental training sample set based on the defect level determination and verification results of step S7, and to incrementally train the core algorithm models of steps S3-S6 to continuously optimize the model parameters. This ensures that the recognition accuracy and detection efficiency of the detection model continuously improve with the number of inspections on the production line. Mini-batch incremental training technology is employed, which trains the model only on newly added valid detection samples, eliminating the need to retrain the entire sample set. This balances the effectiveness of model optimization with lightweight computation, adapting to the real-time requirements of the production line. Furthermore, a model performance evaluation mechanism is used to ensure that the performance of the incrementally trained model is no lower than that of the original model.

[0084] The core input to this step is the defect level judgment and verification dataset DS7={ID, defect type, defect size, defect location, defect level, judgment result validity, processing suggestion} output from step S7, and simultaneously retrieves the original full-domain dataset from step S2 for the corresponding sample. S3 step suspected sample dataset S4 Step Spatial Visual Precision Model Set S5 steps to merge datasets S6 Step Defect Identification Dataset All data is bound to a unique sample ID to ensure the integrity of incremental training samples.

[0085] The incremental training process for the four core algorithms is as follows: Figure 3 As shown, the specific implementation process is as follows: 1. Incremental Training Sample Set Construction: Based on the validity of the judgment results in step S7, valid samples are selected from all detection samples to construct the incremental training sample set. Valid samples include qualified samples with valid judgment results, Level 1 defect samples, Level 2 defect samples, and Level 3 defect samples. Suspected false positive samples and manually re-inspected samples are removed (they will be added back if the re-inspection results are valid after completion). Data annotation and integration are performed on the selected valid samples. Each group of incremental training samples contains complete input data and annotation labels: incremental samples from the improved lightweight convolutional neural network model in step S3, containing global lightweight feature vectors. Labeled with qualified / suspected / clear defects; incremental samples from the S4 step radial basis function implicit surface reconstruction algorithm, including 3D coarse-scan point cloud data. Labels for standard 3D point cloud models; incremental samples from the S5 step tensor product kernel spatial fusion algorithm, containing sensor feature sets. Spatial feature set With pass / defect labels; incremental samples of the S6 step graph attention-enhanced capsule network algorithm, including a multimodal fusion detection model. , final fusion feature tensor Complete labels for defect type, defect size, and defect location. All labeled incremental samples are categorized by model type, constructing incremental training sample sets for four models. Simultaneously, the incremental sample sets are divided into training, validation, and test subsets in a 7:2:1 ratio to ensure uniform sample distribution.

[0086] 2. Incremental Training of Core Algorithm Models: The four core algorithm models from steps S3 to S6 are incrementally trained sequentially. All models employ a mini-batch incremental training mode with a batch size of 16, balancing training efficiency and model stability. The training process is implemented using GPU computing power from the edge computing module. Step S3 improves the lightweight convolutional neural network model by loading the parameters of the original model as initial parameters. The incremental training subset is input into the model, and a stochastic gradient descent optimizer is used for training. The learning rate is set to 0.001, and the learning rate decay coefficient is set to 0.9. The learning rate decays once every 10 training epochs. The cross-entropy loss function is still used. Training continues until the loss function value of the validation subset converges or the number of training epochs reaches 20. Step S4 uses the radial basis function implicit surface reconstruction algorithm, loading the parameter set of the original algorithm as initial parameters. The 3D coarse-scanned point cloud data from the incremental training subset is input into the algorithm, and the support radius of the basis function is adjusted. With weighting coefficients This minimizes the mean square error between the reconstructed model and the standard 3D point cloud model. The S5 step tensor product kernel spatial fusion algorithm loads the parameter set of the original algorithm as initial parameters, inputs the sensor feature set and spatial feature set of the incremental training subset into the algorithm, and uses the Adam optimizer to update the kernel width. With kernel weight matrix The loss function used is cross-entropy loss, and the classification accuracy on the validation subset reaches over 99% after training. The S6 step of the capsule network algorithm with enhanced graph attention loads the parameters of the original model as initial parameters, inputs the fused data of the incremental training subset into the model, and updates the graph attention weight vector. The capsule layer weight matrix and the dynamic routing layer transformation matrix are used. The loss function is a weighted sum of classification loss and localization loss. The defect recognition accuracy of the training to the validation subset reaches more than 99.5% and the localization accuracy reaches 0.01mm.

[0087] 3. Model Performance Evaluation and Parameter Update: After incremental training of each model type, a comprehensive evaluation of model performance is conducted using the corresponding test subset. The evaluation metrics are determined according to the model type: For the improved lightweight convolutional neural network model in step S3, the evaluation metrics are initial screening accuracy and detection speed; for the radial basis function implicit surface reconstruction algorithm in step S4, the evaluation metrics are reconstruction mean square error and reconstruction speed; for the tensor product kernel space fusion algorithm in step S5, the evaluation metrics are defect identification accuracy and fusion speed of fused features; for the graph attention-enhanced capsule network algorithm in step S6, the evaluation metrics are defect identification accuracy, defect localization accuracy, defect size calculation error, and identification speed. If the performance of the incrementally trained model is better than the original model, the incrementally trained model parameters are used as the new optimal parameters, updated and stored in the edge computing module, replacing the original model parameters for subsequent detection; if the performance of the incrementally trained model is not better than the original model, the incremental training is abandoned, the original model parameters are continued to be used, and sample data issues are analyzed. Incremental training is then performed again after accumulating more effective samples.

[0088] 4. End-to-End Algorithm Collaborative Optimization: Based on the incremental training results of a single model, collaborative optimization of the entire algorithm process is performed. The core is to adjust the interaction parameters between steps S4-S6, including the learning rate of step S4. Learning rate of step S5 The weighting coefficients of the loss function in step S6, etc., make the bidirectional interaction of the three algorithms smoother and the closed-loop optimization effect more significant. Through combined testing of multiple sets of interaction parameters, the parameter combination that optimizes the detection accuracy and efficiency of the entire process is selected as the collaborative optimization parameters of the entire process algorithm. These parameters are stored in the edge computing module for the accurate processing of all subsequent suspected samples, realizing the upgrade from single model optimization to the optimization of the entire process algorithm system.

[0089] This step outputs a model optimization parameter set DS8={ID, model type, number of incremental training samples, new model parameters, performance evaluation metrics, and collaborative optimization parameters}. The new model parameters are the optimal parameters updated after incremental training, the performance evaluation metrics are the various detection metrics after model optimization, and the collaborative optimization parameters are the interactive optimization parameters of the entire algorithm. This parameter set is stored in the edge computing module and updated in real time to each detection step for use in subsequent production line inspections. Simultaneously, the incremental training sample set is added to the historical dataset for health product packaging inspection, continuously expanding the historical dataset and providing a richer sample base for subsequent model training.

[0090] S9: Production line collaborative execution and end-to-end feedback of test results This step is the final execution stage of the entire testing process. Its core objective is to achieve collaborative execution with the health supplement production line based on the defect level assessment results from step S7 and the model optimization parameters from step S8. This includes sample sorting, release, and rejection, while simultaneously providing feedback on all test results throughout the entire process, reinforcing steps S1-S8 to form a closed-loop system of testing-execution-feedback-optimization. Production line data interaction and command execution technology is employed, using a data interaction gateway to achieve real-time data transmission between the testing system and the production line control system. This ensures that test results are quickly transformed into production line execution commands, and a test result feedback system is constructed to achieve traceability and full-process feedback of all test data.

[0091] The input data for this step comes from two core sources: first, the defect level judgment and verification dataset DS7={ID, defect type, defect size, defect location, defect level, judgment result validity, and handling suggestions} output from step S7, which serves as the core basis for collaborative execution on the production line; second, the model optimization parameter set DS8={ID, model type, number of incremental training samples, new model parameters, performance evaluation indicators, and collaborative optimization parameters} output from step S8, which serves as the basis for updating detection system parameters and optimizing subsequent detection. All data is bound to a unique sample identification code ID.

[0092] Production line collaborative execution and full-process feedback process, such as Figure 4 As shown, the specific implementation process is as follows: 1. Data interaction between inspection results and the production line control system: The edge computing module establishes a real-time communication connection with the health product production line control system through a data interaction gateway. The communication protocol adopts the industrial Ethernet protocol to ensure the real-time performance and stability of data transmission; the defect level judgment and verification dataset from step S7 is then used to... The unique identification code ID of each sample is transmitted to the production line control system in real time. Based on the processing suggestions in the dataset, the production line control system generates corresponding execution instructions for each sample. The execution instructions correspond one-to-one with the execution mechanisms of the production line, including conveyor, sorting robot, rejection robot, warning indicator, etc.

[0093] 2. Production Line Collaborative Execution: Based on generated execution instructions, the production line control system controls each actuator to complete corresponding operations. Operations and processing suggestions strictly correspond to ensure accurate sample sorting and processing. For qualified samples marked "Direct Release," the control system maintains a constant speed on the conveyor, transporting the sample to the next production process without further operation. For first-level defect samples marked "Release After Warning," the control system activates a yellow warning indicator and simultaneously controls the conveyor to operate normally, transporting the sample to the next process. The warning information is simultaneously transmitted to the production line central control system, recording the sample's ID and defect information. For second-level defect samples marked "Re-inspection After Sorting," the control system controls the sorting robot to move, separating the sample from the main conveyor to the re-inspection conveyor. Samples sent to the re-inspection station on the production line undergo secondary testing by the re-inspection equipment, and the re-inspection results are re-entered into the testing system. For Level 3 defect samples marked "directly rejected," the control system controls the rejection robot to remove the sample from the main conveyor to the defective product collection box, while recording information such as the defect sample's ID, defect type, and defect level for subsequent quality analysis. For suspected false positive samples marked "secondary inspection," the control system controls the conveyor to transport the sample to the secondary inspection station in the testing area, restarting the testing process of steps S2-S7. For samples marked "manual re-inspection," the control system controls the sorting robot to sort the sample to the manual re-inspection station, where professional quality inspectors perform manual inspection. The manual inspection results are manually entered into the testing system, marked as qualified / defective, and the specific defect information is recorded.

[0094] 3. Full-process feedback of test results: Construct a full-process feedback system for test results, integrate the test results of all samples, production line execution results, and manual re-inspection results according to the sample's unique identification code ID, and form a complete test file. The test file contains all information such as sample production information, collection data, initial screening results, precision processing results, defect identification results, grade determination results, execution operation results, and re-inspection results, so as to achieve full life cycle traceability of test data. The integrated detection results are fed back to steps S1-S8 according to different uses: Step S1 optimizes the calibration cycle and parameters of the hardware module; if the detection deviation of a certain type of sensor continues to increase, the calibration cycle of that sensor is shortened; Step S2 optimizes the triggering timing and acquisition parameters of data acquisition to improve the quality of the acquired data; Steps S3-S6 expand the training sample set of each algorithm model to provide more effective samples for subsequent incremental training; Step S7 optimizes the defect level classification standard by adjusting the weights and thresholds of each dimension based on the actual execution and quality feedback of the production line; Step S8 optimizes the incremental training strategy of the model by adjusting parameters such as batch size and learning rate of incremental training based on the distribution of detection results to improve the efficiency of model optimization.

[0095] 4. System Status Monitoring and Maintenance Early Warning: Based on the full-process testing data and execution results, the system monitors the operating status of the testing system in real time. Monitoring indicators include the operating status of hardware modules, data acquisition success rate, algorithm model detection accuracy, and production line actuator action accuracy. If a monitoring indicator exceeds the preset normal range, such as the hardware module acquisition success rate being lower than 99% or the algorithm model detection accuracy being lower than 99%, the testing system automatically generates maintenance early warning information. The early warning information includes abnormal indicators, abnormal cause analysis, and maintenance suggestions, and is synchronously transmitted to the production line central control system and equipment maintenance department to remind maintenance personnel to perform timely equipment inspection and maintenance, ensuring the stable operation of the testing system and avoiding testing errors and production line interruptions caused by system failures.

[0096] This step outputs a production line execution and full-process feedback dataset DS9={ID, Sample Production Information, Full-Process Inspection Information, Production Line Execution Result, Re-inspection Result, Inspection Archive, System Operating Status, Maintenance Warning Information}. The inspection archive is a complete inspection record of the sample, enabling full-process traceability. The system operating status is a real-time monitoring indicator of the inspection system. The maintenance warning information is an alert when the system malfunctions. This dataset is stored in the health product production quality traceability system for subsequent product quality queries and analysis. Simultaneously, it provides real-time feedback to steps S1-S8, achieving closed-loop optimization of the entire inspection process. This ensures that the inspection system's accuracy, efficiency, and stability continuously improve with the number of inspections, ultimately achieving intelligent, high-precision, and real-time inspection of the integrity of health product packaging.

Claims

1. A method for detecting the integrity of health product packaging based on collaborative efficiency and intelligent detection, characterized in that, include: S1: Perform spatiotemporal synchronization and accuracy calibration on the hardware module of the detection system to obtain the calibration parameter set of the detection system; S2: Based on the calibration parameter set of the detection system, perform full-domain spatiotemporal synchronous data acquisition on health product packaging samples and bind the sample's unique identification code to obtain the full-domain raw dataset; S3: An improved lightweight convolutional neural network algorithm is used to extract features and perform initial screening of samples from the original dataset across the entire domain to obtain a suspected sample dataset; S4: The radial basis function implicit surface reconstruction algorithm is used to preprocess the point cloud and accurately reconstruct the 3D point cloud of the suspected sample dataset to obtain a set of accurate spatial vision models. S5: The tensor product kernel spatial fusion algorithm is used to perform deep multimodal data fusion by combining the sensor feature set extracted from the suspected sample dataset with the spatial feature set extracted from the spatial vision precision model set to obtain the fused dataset; S6: The capsule network algorithm with graph attention enhancement is used to accurately identify defect features in the fused dataset, resulting in a defect identification dataset; S7: Based on the defect identification dataset, perform defect level determination and full-dimensional cross-validation to obtain the defect level determination and validation dataset.

2. The method for detecting the integrity of health product packaging based on collaborative efficiency and intelligent detection according to claim 1, characterized in that, Also includes: S8: Based on the defect level judgment and verification dataset, an incremental training sample set is constructed to perform incremental training and parameter optimization on the improved lightweight convolutional neural network algorithm, radial basis function implicit surface reconstruction algorithm, tensor product kernel space fusion algorithm and graph attention enhanced capsule network algorithm to obtain the model optimization parameter set; S9: Based on the defect level judgment and verification dataset and the model optimization parameter set, the detection system and the production line control system are coordinated to execute, and the detection results of the whole process are fed back to each step from S1 to S8 to form a closed loop of the whole process of health product packaging integrity detection.

3. The method for detecting the integrity of health product packaging based on collaborative efficiency and intelligent detection according to claim 1, characterized in that, S5 generates spatial registration error during the deep fusion of multimodal data. This spatial registration error is fed back to S4 to adjust the parameters of the radial basis function implicit surface reconstruction algorithm. S6 generates defect feature gradient during the accurate identification of defect features. This defect feature gradient is fed back to S5 to adjust the parameters of the tensor product kernel spatial fusion algorithm.

4. The method for detecting the integrity of health product packaging based on collaborative efficiency and intelligent detection according to claim 3, characterized in that, The specific process of S5 feeding back the spatial registration error to S4 is as follows: S5 calculates the spatial registration error of the two types of features based on the sensor feature set and the spatial feature set. This spatial registration error is transmitted to the radial basis function implicit surface reconstruction algorithm of S4 in real time. Based on this spatial registration error, S4 adjusts the basis function weight coefficients and support radius of its own algorithm. The adjusted algorithm reconstructs the three-dimensional point cloud accurately on the suspected sample dataset, generates an optimized spatial visual accurate model set, and inputs it into S5 to achieve dynamic optimization of reconstruction accuracy.

5. The method for detecting the integrity of health product packaging based on collaborative efficiency and intelligent detection according to claim 3, characterized in that, The specific process of S6 feeding back the defect feature gradient to S5 is as follows: When S6 performs defect feature recognition, it generates a set of digital capsule vectors through the capsule layer operation of the graph attention-enhanced capsule network algorithm. At the same time, it calculates the first-order partial derivative of the total model loss with respect to the set of digital capsule vectors to obtain the defect feature gradient. This defect feature gradient is transmitted to the tensor product kernel spatial fusion algorithm of S5 in real time. Based on this defect feature gradient, S5 adjusts the kernel weight matrix of its own algorithm. The adjusted algorithm recombines the sensor feature set and the spatial feature set to perform multimodal data deep fusion, generates an optimized fusion dataset, and inputs it into S6 to improve the defect recognition accuracy of the fusion features.

6. The method for detecting the integrity of health product packaging based on collaborative efficiency and intelligent detection according to claim 3, characterized in that, S4, S5, and S6 form a two-way interactive closed loop. The termination condition of the closed loop is: when the spatial registration error generated by S5 drops to the preset matching threshold, and the defect identification dataset output by S6 is verified as valid by the full-dimensional cross-validation of S7, the two-way interactive closed loop terminates. S4, S5, and S6 respectively output the final spatial visual accurate model set, fusion dataset, and defect identification dataset, which are used for defect level determination, algorithm incremental training, and production line collaborative execution.

7. The method for detecting the integrity of health product packaging based on collaborative efficiency and intelligent detection according to claim 1, characterized in that, The specific execution process of the improved lightweight convolutional neural network algorithm in S3 is as follows: spatial visual features and sensor physical features are extracted from the original dataset. The two types of features are concatenated to form a lightweight feature vector. This lightweight feature vector is input into the trained improved lightweight convolutional neural network model. The model outputs the category prediction probability of the sample through feature operation. Based on the preset probability threshold, samples that cannot be clearly judged as qualified or clearly defective are selected. All relevant data of such samples are integrated to form a suspected sample dataset.

8. The method for detecting the integrity of health product packaging based on collaborative efficiency and intelligent detection according to claim 1, characterized in that, The specific execution process of the implicit surface reconstruction algorithm of radial basis function in S4 is as follows: 3D coarse-scan point cloud data is extracted from the suspected sample dataset. First, the 3D coarse-scan point cloud data is preprocessed by removing outliers and extracting key points of potential defect areas. Then, based on the preprocessed point cloud data, a compactly supported radial basis function is constructed and the implicit surface model parameters are solved. Subsequently, a laser 3D scanner is controlled to perform high-precision 3D fine scanning on the suspected samples to obtain fine-scan point cloud data. The fine-scan point cloud data is fitted to the implicit surface model to complete the 3D point cloud reconstruction. Finally, the spatial features are bound together with the six-view visual features of the suspected samples and the reconstructed point cloud model to generate a precise spatial visual model set.

9. The method for detecting the integrity of health product packaging based on collaborative efficiency and intelligent detection according to claim 1, characterized in that, The full-dimensional cross-validation in S7 includes three independent validation levels: sensor data validation, spatial vision data validation, and fusion feature validation. Among them, sensor data verification verifies the physical matching of defect features based on sensor feature sets extracted from suspected sample datasets, spatial vision data verification verifies the spatial matching of defect features based on spatial vision precision model sets, and fusion feature verification verifies the fusion matching of defect features based on fusion datasets. When all three levels of verification results pass, the defect identification dataset is deemed valid, and a corresponding defect level judgment and verification dataset is generated based on the defect identification dataset. If any level of verification fails, the relevant data of the corresponding sample will be re-included in the suspected sample dataset, and a secondary detection process will be initiated.

10. The method for detecting the integrity of health product packaging based on collaborative efficiency and intelligent detection according to claim 2, characterized in that, The construction and incremental training process of the incremental training sample set in S8 is as follows: Select samples with valid judgment results from the defect level judgment and verification dataset, remove suspected false detection samples, classify and integrate the relevant data of the valid samples according to the algorithm type, and construct and improve the incremental training sample sets corresponding to the lightweight convolutional neural network algorithm, the radial basis function implicit surface reconstruction algorithm, the tensor product kernel space fusion algorithm, and the graph attention-enhanced capsule network algorithm. Perform small batch incremental training for the four types of algorithms and adjust the core parameters of each algorithm. After training, integrate the optimization parameters of all algorithms to form the model optimization parameter set.

Images